Timeline des Sorties de Modèles IA

2.8 trillion parameter mixture-of-experts model with 1 million token context window and native multimodal capabilities

Introduces Kimi Delta Attention, enabling up to 6.3x faster decoding in million-token contexts

Attention Residuals technique delivers ~25% higher training efficiency at less than 2% additional cost

Built for long-horizon agentic coding and self-evolving workflows

Demonstrates frontier-level performance across evaluations, outperforming other tested open models while trailing Claude Fable 5 and GPT 5.6 Sol

Live on Kimi.com, Kimi Work, Kimi Code, and the Kimi API

Open weights release planned by July 27, 2026

Official Kimi API achieves above 90% cache hit rate in coding workloads via Mooncake disaggregated inference architecture

Grok 4.5 is SpaceXAI frontier model built for coding, agentic tasks, and knowledge work.

Trained in SpaceXAI Memphis data centers on new datasets spanning science, engineering, and math.

Served at 80 TPS (tokens per second), faster than flash-class models.

Delivers 4.2x fewer output tokens than Claude Opus 4.8 (max) on SWE Bench Pro (15,954 vs 67,020 avg tokens).

Scores 83.3% on Terminal-Bench 2.1 vs 78.9% for Claude Opus 4.8.

Reaches 62% on DeepSWE 1.0 vs 55.8% for Claude Opus 4.8.

Claude Opus 4.8 still leads on SWE-Bench Multilingual (84.4% vs Grok 4.5 78%).

Claude Opus 4.8 also leads SWE-Bench Pro (69.2% vs Grok 4.5 64.7%).

Supports reasoning modes: low, medium, or high (default high).

Available via Responses API and Chat Completions, with function calling, web search, X search, and code execution tools.

Not yet available in the EU; EU availability expected mid-July.

Cursor participated in training and published benchmark results.

295B parameter Mixture-of-Experts (MoE) architecture

Best in its size class, rivaling trillion-scale flagship models

Apache 2.0 license, friendly for commercial use

Released as free tier on OpenRouter as tencent/hy3:free

Following the late-April preview, Tencent gathered feedback from 50+ products and scaled up post-training with higher-quality data

Blind evaluation with 270 domain experts scored Hy3 at 2.67/4, outperforming GLM-5.1 at 2.51/4

Largest advantages seen in frontend development, data & storage, and CI/CD tasks

Hallucination rate dropped from 12.5% to 5.4%

Commonsense error rates fell from 25.4% to 12.7%

Used 47.4% fewer tokens than GLM-5.2 for document processing

Used 49% fewer tokens than GLM-5.2 for presentation creation

Reliable and affordable for most agentic use cases

33B total / 3B activated parameters Mixture-of-Experts model for agentic coding

SWE-bench Multilingual up 5.4 points to 63.1% versus XS.2

Same architecture as XS.2; gains come from a training and data refresh

OpenMDW-1.1 license — fully permissive, aligned with NVIDIA and Linux Foundation direction

DFlash open-weighted speculative decoders roughly double achieved tokens per second

256K context length served on Poolside API and OpenRouter

Quantized checkpoints available: FP8, INT4, NVFP4 (GGUF coming with llama.cpp)

Supported in vLLM, SGLang, NVIDIA TensorRT-LLM, HuggingFace Transformers, and Ollama

XS.2 will be sunset on Poolside API after 1 week; remains available via Baseten Model Library

1.6T parameter MoE model with ~48B active parameters and 1M token context window

The full model behind Owl Alpha available on OpenRouter, now released openly

LongCat Sparse Attention (LSA) scales efficiently for 1M-context tokens

Zero-Compute Experts dynamically activate 33B-56B parameters per token with zero wasted compute

MOPD architecture uses three specialized expert groups (Agent / Reasoning / Interaction), gate-routed per task

Scores 70.8 on Terminal-Bench 2.1

Scores 59.5 on SWE-bench Pro, beating GPT-5.5 at 58.6

Scores 77.3 on SWE-bench Multilingual

Scores 73.2 on FORTE, 78.8 on RWSearch, and 79.9 on BrowseComp

LFM2.5-230M is Liquid AI's smallest model yet, built to run fast anywhere (CPUs, NPUs, and GPUs) to enable agentic tasks on phones, robots, home and network automation devices.

It has 230M parameters and is built on the LFM2 architecture.

Pre-trained on 19T tokens, with a 32K context extension.

Post-trained with distillation from LFM2.5-350M.

Achieves 213 tok/s decode speed on a Samsung Galaxy S25 Ultra (CPU) and 42 tok/s on a Raspberry Pi 5 (CPU).

Competes with and often beats models more than twice its size on instruction following, data extraction, and tool use.

On Raspberry Pi 5 and Qualcomm Snapdragon Gen4 (Galaxy S25 Ultra), it delivers the highest prefill and decode throughput in its class while keeping the smallest memory footprint.

Available on all platforms: llama.cpp (GGUF) for edge, MLX for Apple Silicon, vLLM and SGLang for GPU serving, and ONNX for cross-platform deployments.

In an early demo, LFM2.5-230M was deployed on a Unitree G1 robot running entirely on-device on its onboard Jetson Orin, acting as a skill-selection layer that decomposes natural-language instructions into structured multi-step tool-call plans.

For production-grade enterprise deployments, Liquid AI developed an internal GPU inference stack that delivers extremely low-latency serving, with LFM2.5-230M achieving considerably lower end-to-end latency than other small models on SGLang across all concurrency levels.

Well-suited for large-scale data extraction pipelines and lightweight on-device agentic workloads.

LFM2.5-230M and LFM2.5-230M-Base are available now.

Ornith-1.0 is a family of agentic coding models spanning four parameter sizes: 9B Dense, 31B Dense, 35B MoE, and 397B MoE.

It achieves state-of-the-art performance among open-source models of comparable size on coding benchmarks: Terminal-Bench 2.1 (77.5), SWE-Bench Verified (82.4), SWE-Bench Pro (62.2), SWE-Bench Multilingual (78.9), NL2Repo (48.2), SWE Atlas (41.2 QnA / 42.6 RF / 39.1 TW), and ClawEval (77.1).

The models are post-trained on top of Gemma 4 and Qwen 3.5 base models.

Ornith-1.0 employs a novel self-improving training strategy in which reinforcement learning is used to generate not only solution rollouts, but also the task-specific scaffolds that drive those rollouts, jointly optimizing scaffold and solution for higher-quality agentic coding.

Ornith-1.0-397B (77.5 on Terminal-Bench 2.1, 82.4 on SWE-Bench Verified) matches the performance of Claude Opus 4.7 (70.3 on TB-2.1, 80.8 on SWE-Bench Verified).

Ornith-1.0-397B outperforms leading open-source models of similar size including MiniMax M3 (66.0 TB-2.1, 80.5 SWE-Bench Verified) and DeepSeek-V4-Pro (67.9 TB-2.1, 80.6 SWE-Bench Verified).

Ornith-1.0-9B, deployable on edge devices, matches or exceeds the performance of much larger models such as Gemma 4-31B and Qwen 3.6 35B.

All models are released under the MIT license, enabling full commercial and research use.

Sakana AI launched Sakana Fugu on 22 June 2026 as a multi-agent orchestration system that behaves like a single model behind one OpenAI-compatible API.

Fugu is itself a language model trained to call other LLMs in an agent pool, including instances of itself called recursively; it handles model selection, delegation, verification and synthesis internally.

Instead of hard-coded roles, Fugu learns how to coordinate — deciding when to delegate, how agents communicate, and how to combine their outputs into one answer.

Ships in two variants: Fugu (balanced performance/low latency, everyday coding, review, chatbots, Codex-compatible) and Fugu Ultra (max quality on hard multi-step problems).

Fugu allows opting specific agents out of its pool for data, privacy and compliance needs; Fugu Ultra has a fixed pool with no opt-out.

Current Fugu Ultra model ID is fugu-ultra-20260615.

Fugu Ultra leads most published coding and reasoning benchmarks, and the orchestrator beats the individual models it coordinates.

Sakana frames Fugu as a hedge against single-vendor dependency — if a provider restricts access (the team cites recent export controls on Anthropic Fable and Mythos models), Fugu routes around it.

Routing is proprietary: per-query model selection stays hidden from the caller.

Subscription tiers (monthly, both Fugu and Fugu Ultra included): Standard $20/month, Pro $100/month (10x Standard), Max $200/month (30x Standard); subscribe before end of July 2026 for a free second month.

Flagship foundation model with truly usable 1M-token context window for project-scale engineering

128K maximum output tokens for long-form generation

Open-source under MIT license with weights available on HuggingFace and ModelScope

Introduces IndexShare architecture reducing per-token FLOPs by 2.9x at 1M context length

Improved speculative decoding with 20% acceptance length increase via MTP with IndexShare and KVShare

Highest-ranked open-source model on FrontierSWE (74.4%), trailing Claude Opus 4.8 by only 1%

Strongest open-source coding model: 81.0 on Terminal-Bench 2.1, 62.1 on SWE-bench Pro

Supports multiple thinking effort levels (High and Max) to balance performance and latency

Achieves 99.2% on AIME 2026 and 91.2% on GPQA-Diamond reasoning benchmarks

Introduces anti-hack module for coding RL training to prevent reward hacking

Supports function calling, context caching, structured output, streaming, and MCP integration

+21.8% improvement on Kimi Code Bench v2 vs K2.6

+11% improvement on Program Bench vs K2.6

+31.5% improvement on MLS Bench Lite vs K2.6

30% fewer tokens on internal reasoning vs K2.6

256k context window with long thinking and deep reasoning support

Native multimodal architecture supporting text, image, and video input

Supports ToolCalls, JSON Mode, Partial Mode, and automatic context caching

Open-source weights available on HuggingFace

Available via API and in Kimi Code IDE

Separate 6x High-Speed mode coming soon

Beta program open for early access to future updates

North Mini Code is a code generation and reasoning model released by Cohere on 9 June 2026.

Released under the Apache 2.0 license as open source, allowing self-hosting and commercial use.

Supports a 256K token context window with up to 64K tokens of output, suited to large repositories and long technical documents.

Text-in, text-out only (no multimodal inputs or outputs).

Scores 75.7% on GPQA Diamond, placing it among advanced generalist reasoning models.

Additional benchmarks: SciCode 38.2%, TAU2-bench 37.4%, TerminalBench-Hard 31.1%, LiveCodeBench Reasoning 32.3%, IF-Bench 57.6%.

Originally listed at $0/M tokens; Cohere has since moved the model to paid pricing — check the official Cohere pricing page for current rates.

Well suited as a local code assistant via Ollama or llama.cpp for audits, test generation, refactors and automated PR review on large monorepos.

Claude Fable 5 is a Mythos-class model made safe for general use

State-of-the-art on nearly all tested AI capability benchmarks

Exceptional performance in software engineering, knowledge work, vision, and scientific research

The longer and more complex the task, the larger Fable 5 lead over other Anthropic models

Safeguards route some cybersecurity queries to Claude Opus 4.8 instead; triggers in less than 5% of sessions on average

Claude Mythos 5 is the same underlying model with safeguards lifted in some areas

Mythos 5 deployed through Project Glasswing with the US government as an upgrade to Claude Mythos Preview

Mythos 5 has the strongest cybersecurity capabilities of any model in the world

Access to Mythos 5 will expand through a broader trusted access program

Pricing is $10 per million input tokens and $50 per million output tokens — less than half the price of Claude Mythos Preview

Available via the Claude API as claude-fable-5

749B-class Mixture-of-LoRA (MoL) agent model: 744B frozen base (GLM-5.1) + 5 specialist LoRA adapters (~1B each)

Post-trained from GLM-5.1 using MinT post-training infrastructure

Five specialist LoRAs: L0 (default chat/general), L1 (personal-life tasks), L2 (coding), L3 (A2UI Generative UI), L4 (OpenClaw-style agent tasks)

Router Tool design exposes model selection as a standard tool call, routing between LoRAs via an explicit change_model tool — debuggable and compatible with vLLM OpenAI server mode

202,752 token context window, bfloat16 precision

Released under MIT open-source license on Hugging Face as a single repository (base at root, LoRAs under l0/ through l4/)

Co-designed with production Agent Harness via Harness Context Protocol (HCP) — same routing, memory, and tool-call tokenization during training and serving

R3 (Rollout Routing Replay) for provable expert-path alignment during MoE RL training, combined with IcePop-style rollout correction and DSA attention alignment fixes

Self-evolution capability through AutoResearch + Context Learning loop: model improves its own prompts and scaffolds, then distills improved trajectories back into parameters

Macaron LivingBench: in-house dynamic personal-life agent benchmark with coupled dynamic noise, dynamic environment, and dynamic user simulation

Trained on A2UI protocol for Generative UI with 3ms TPOT latency in interactive scenarios via TileRT collaboration

A2UI-Bench evaluates protocol correctness, task construction correctness, and real user-experience lift with visual-side evaluation

MindForge agentic RL training framework brings the production harness directly into the RL loop

Live preview available at macaron.im; 30B and 200B open-source variants planned for V1 non-preview release

Managed inference and post-training coming soon on MinT platform

550B total parameters with 55B active parameters using Mixture-of-Experts architecture

Hybrid Mamba-Attention architecture with LatentMoE for improved expert routing

Multi-Token Prediction (MTP) layers for faster inference through native speculative decoding

Pretrained in NVFP4 precision, running on Hopper, Blackwell, and Ampere GPUs with a single checkpoint

Post-trained with SFT, Reinforcement Learning, and Multi-Teacher On-Policy Distillation (MOPD) using 10+ specialized teacher models

Achieves 5.9x, 4.8x, and 1.6x higher inference throughput vs GLM-5.1, Kimi-K2.6, and Qwen-3.5 respectively on 8k/64k token setting

Supports context length up to 1M tokens, outperforming state-of-the-art open LLMs on RULER at 1M context

Lowers cost of complex agentic tasks by up to 30% while delivering frontier accuracy

Fully open: weights, training data (173B code tokens, legal data, specialized data), and training recipes all released

Licensed under OpenMDW 1.1, the Linux Foundation permissive license for open AI model distributions

Available in multiple checkpoints: NVFP4, BF16, Base BF16, and GenRM

Deployable on-premise, on-cloud, or at-the-edge via NVIDIA NIM and major providers

A unified, encoder-free multimodal model designed to bring high-performance intelligence directly to your laptop, released under an Apache 2.0 license.

Laptop Ready: Small enough to run locally with just 16GB of VRAM or unified memory.

Unified Architecture: Multimodal tokens flow directly into the LLM backbone. No additional encoders are needed.

Advanced Reasoning: Gemma 4 12B delivers benchmark performance nearing the larger 26B model, but at less than half the memory footprint. Unlocks powerful multi-step reasoning and agentic workflows.

Vision Embedding: The vision encoder is replaced with a lightweight 35M-parameter module. By injecting spatial information directly into the token embeddings, the unified model takes over visual understanding.

Broad Ecosystem Support: Weights available on Hugging Face and Kaggle, compatible with llama.cpp, MLX, LM Studio, vLLM, and SGLang.

Bridges the gap between edge efficiency and advanced reasoning, making it the best model for self/local hardware on a low budget.

MoE (Mixture of Experts) architecture with 397B total parameters and 17B active parameters

Post-trained on Qwen3.5-397B-A17B base model

262K token context window with up to 256K output tokens

Accepts text and image input, produces text output

Released under Apache-2.0 open-source license

Scores 75.3 on Terminal-Bench 2.1 for coding tasks

Scores 1585 on GDPval for long-running workflows

Achieves SOTA among open-source models on SWE-Verified, SWE-Pro, and DeepSWE benchmarks

Agentic Thinking capability: unifies reasoning, tool use, and environmental execution in a closed loop (comprehension → planning → implementation → feedback → debug → iteration)

Adaptive reasoning depth reduces thinking tokens by 30-50% compared to always-on reasoning

Available on Hugging Face, ModelScope, SiliconFlow (early access serverless), and OpenRouter

Can run locally via llama.cpp, Ollama, and similar tools

Native integration with Claude Code, Cursor, OpenClaw, and other agentic harnesses

Rivals GPT-5.5 and Opus 4.7 performance, reaching top-tier level

Also available as Nex-N2-mini variant

Multimodal interactive hybrid agent: unified GUI & CLI operation across visual and text tasks

Versatile coding agent & productivity assistant with full-modality input

Visual Agent: perception, reasoning, grounding, and search-augmented QA

Cross-harness generalization across diverse agent frameworks

Multimodal improvements extend beyond isolated visual understanding gains

Achieves top-tier performance on coding and agentic benchmarks with autonomous task decomposition, tool use, and multi-step reasoning capabilities

Powered by proprietary MiniMax Sparse Attention (MSA) architecture supporting 1M token context window with guaranteed minimum of 512K tokens

Natively multimodal model

On BrowseComp, scores 83.5 surpassing Opus 4.7 (79.3)

First open model to simultaneously achieve frontier coding capabilities, million-token context, and multimodal support

API pricing: input $0.60/M tokens (≤ 512k), $1.20/M tokens (> 512k); output $2.40/M tokens (≤ 512k), $4.80/M tokens (> 512k); prompt caching $0.12/M tokens (≤ 512k), $0.24/M tokens (> 512k)

#1 on ClawEval-1.1 with a score of 67.1

#1 on SimpleVQA Search with a score of 79.2

#2 on SWE-PRO with a score of 56.3

95.3 on V* Python benchmark

400 tokens per second throughput

198B total parameters, ~11B active (sparse MoE architecture)

256K context window with 3 reasoning levels

Native multimodal — understands UIs, charts, documents, and images, then writes code or calls tools

Web + visual search with more sources and deeper follow-up

98%+ on τ²-bench across all difficulty levels for reliable tool use

Open weights released under Apache 2.0 license

Builds on Opus 4.7 with stronger performance across coding, agentic tasks, and professional work

Around 4x less likely than its predecessor to allow flaws in code it has written to pass unremarked

Higher honesty — more likely to flag uncertainties and less likely to make unsupported claims

Only model to complete every case end-to-end on the Super-Agent benchmark, beating prior Opus models and GPT-5.5 at parity on cost

Scored 84% on Online-Mind2Web, making it the strongest computer-use and browser-agent model tested

Highest score recorded on the Legal Agent Benchmark and first model to break 10% on the all-pass standard

Rates of misaligned behavior substantially lower than Opus 4.7, similar to Claude Mythos Preview

Launches alongside dynamic workflows in Claude Code for running hundreds of parallel subagents

New effort control feature in claude.ai lets users choose how much effort Claude puts into a response

Fast mode runs at 2.5x speed and is now 3x cheaper than for previous Opus models

API model ID is claude-opus-4-8; Messages API now accepts system entries inside the messages array

Agentic flagship model designed for autonomous agents capable of coding, orchestrating workflows via MCP, and sustaining long action chains on multi-step tasks

35-hour autonomous kernel optimization test with over 1,000 tool calls without breaking the reasoning chain

Good cross-framework generalization including Claude Code, OpenClaw, and Qwen Code

Currently offers a pure text-only interface for public experimentation

Deep thinking capabilities for complex reasoning tasks

Deployed via Aliyun Bailian API at $2.5 input / $7.5 output per million tokens

Google high-efficiency multimodal model delivering near-Pro level coding and reasoning at Flash-tier cost and speed

Supports text, image, video, audio, and PDF inputs natively

Defaults to medium thinking effort with full support for thinking levels: minimal, low, medium, high for fine-grained cost/performance tuning

Surpasses Gemini 3.1 Pro on coding and agentic benchmarks: Terminal-Bench 2.1 (76.2%), GDPval-AA (1656 Elo), MCP Atlas (83.6%)

Leads multimodal understanding with 84.2% on CharXiv Reasoning benchmark

4x faster output tokens per second compared to other frontier models

Ranked in the upper-right quadrant of the Artificial Analysis Intelligence Index for top intelligence at exceptional speed

Ideal for long-horizon agentic tasks at less than half the cost of competing frontier models

Integrates with Antigravity for collaborative sub-agent deployment at enterprise scale

1M token context window

Tops LMArena Search leaderboard as #4 globally and #1 among Chinese models with a score of 1,223

Math: #9 globally on LMArena category leaderboards

Legal & Government: #1 globally on LMArena category leaderboards

Business, Management & Financial Ops: #4 globally on LMArena category leaderboards

Software & IT Services: #7 globally on LMArena category leaderboards

Compresses total parameters to approximately 1/3 and active parameters to approximately 1/2 of ERNIE-5.0

Achieves leading performance using only about 6% of the pre-training cost of comparable models

Powered by decoupled fully-asynchronous reinforcement learning infrastructure built on PaddlePaddle

Uses scaled agentic post-training with Multi-Teacher On-Policy Distillation (MOPD)

Scores 99.6 on AIME26 with tool use, second only to Gemini 3.1 Pro

Surpasses DeepSeek-V4-Pro on tau-cubed-bench and SpreadsheetBench-Verified agent evaluation tasks

Four-stage post-training pipeline: SFT, Domain Expert Model Training, On-Policy Distillation, and General Online RL

Based on Once-For-All elastic training framework with elastic depth, width, and sparsity dimensions

Rolling out on 10+ creative production agent platforms including ISEKAI ZERO and Mulan AI

Creative writing capabilities approach those of Gemini 3.1 Pro

Reasoning model from xAI with configurable effort levels (none/low/medium/high, default low)

Accepts text and image inputs with text output

Suited for agentic workflows, instruction-following tasks, and high factual accuracy applications

1 million token context window with no output token limit

Well-suited for long-document analysis, deep research, and multi-step agentic tasks

Tiered pricing: requests exceeding 200k total tokens are billed at a higher rate

New flagship model merging instruction-following, reasoning, and coding into a single 128B dense architecture

Released as open weights under a modified MIT license

Runs self-hosted on as few as four GPUs

API pricing at $1.50/mtok input and $7.50/mtok output

Powers the new Mistral Vibe remote agents for async cloud coding sessions

Drives Work mode in Le Chat for multi-step agentic task execution with parallel tool calling

Sessions can be spawned from CLI or Le Chat, and local CLI sessions can be teleported to the cloud

Multimodal model unifying video, audio, image, and text understanding in a single architecture

Hybrid Mixture-of-Experts (MoE) 30B-A3B architecture with 30B total and 3B active parameters

Up to 9x higher throughput compared to similar open omnimodal models

256K unified context window with single-pass perception

Hybrid architecture combining Mamba layers for memory efficiency and transformers for precise reasoning

Integrates vision encoders (C3D for video) and audio encoders (Paraquet), eliminating need for separate models

Supports FP8/NVFP4 quantization with optimized inference on NVIDIA Ampere, Hopper, and Blackwell GPUs

Designed for enterprise multimodal agents: document intelligence (OCR, tables), GUI navigation, audio-video reasoning

Runs locally with 25-36GB RAM in 4/8-bit quantization via Unsloth or vLLM

Available on Hugging Face, Ollama, OpenRouter, and NVIDIA NIM

225B total parameter Mixture-of-Experts model with 23B activated parameters per token

Poolside most capable model to date, completed pre-training at end of 2025

Trained from scratch on 30T tokens using Muon optimizer

Trained on 6,144 interconnected NVIDIA Hopper GPUs entirely in-house

Achieves 72.5% on SWE-bench Verified, 67.3% on SWE-bench Multilingual, 46.9% on SWE-bench Pro, 40.7% on Terminal-Bench 2.0

128K context window with up to 8K output tokens

Agentic coding model built for long-horizon software engineering tasks

Foundation for the entire Laguna model family

Uses custom async on-policy RL system with Agent Client Protocol (ACP) server

Free to use for a limited time via poolside API and OpenRouter

Weights available on request for startups, institutions, and universities

33B total parameter Mixture-of-Experts model with 3B activated parameters per token

First open-weight release from poolside, licensed under Apache 2.0

Trained on 30T tokens using Muon optimizer

Supports native reasoning with interleaved thinking between tool calls

Uses Sliding Window Attention with per-head gating in 30 of 40 layers

KV cache quantized to FP8 for reduced memory per token

Compact enough to run locally on a Mac with 36 GB RAM

128K context window with up to 8K output tokens

Achieves 68.2% on SWE-bench Verified, 62.4% on SWE-bench Multilingual, 44.5% on SWE-bench Pro, 30.1% on Terminal-Bench 2.0

Supports vLLM, Transformers, TRT-LLM, and Ollama

Agentic coding model built for long-horizon software engineering tasks

Free to use for a limited time via poolside API and OpenRouter

Deux modèles : DeepSeek-V4-Pro (1.6T total / 49B active params) et DeepSeek-V4-Flash (284B total / 13B active params)

Context length de 1M tokens, output max de 384K tokens

Support thinking mode (par défaut) et non-thinking mode

Pricing ultra-agressif : Flash à $0.14/M input tokens (cache miss), $0.028/M (cache hit), $0.28/M output — soit ~7x moins cher que Claude Opus 4.7

Pro à $1.74/M input tokens (cache miss), $0.145/M (cache hit), $3.48/M output

Modèles open-source, poids disponibles sur HuggingFace

Compatible format API OpenAI et Anthropic (https://api.deepseek.com et https://api.deepseek.com/anthropic)

Support JSON output, Tool Calls, Chat Prefix Completion (Beta), FIM Completion (Beta)

Performance rivalisant avec les meilleurs modèles closed-source mondiaux

GPT-5.5 is OpenAI smartest and most intuitive to use model yet, described as the next step toward a new way of getting work done on a computer

Achieves 82.7% on Terminal-Bench 2.0, 73.1% on Expert-SWE (Internal), and 84.9% on GDPval — all state-of-the-art scores

Matches GPT-5.4 per-token latency while performing at a much higher level of intelligence

Significantly more token efficient — uses fewer tokens to complete the same tasks compared to GPT-5.4

Scores 78.7% on OSWorld-Verified for real computer environment operation and 81.8% on CyberGym

GPT-5.5 Pro achieves 90.1% on BrowseComp and 52.4% on FrontierMath Tier 1-3

On SWE-Bench Pro, reaches 58.6% solving more tasks end-to-end in a single pass than previous models

Proactively deployed with industry-leading cybersecurity safeguards, classified as High under OpenAI Preparedness Framework

Helped discover a new proof about Ramsey numbers in combinatorics, later verified in Lean

Scores 25.0% on GeneBench for multi-stage scientific data analysis in genetics

API pricing: $5/1M input tokens and $30/1M output tokens with 1M context window

GPT-5.5 Pro API pricing: $30/1M input tokens and $180/1M output tokens

Co-designed, trained with, and served on NVIDIA GB200 and GB300 NVL72 systems

Rolling out to Plus, Pro, Business, and Enterprise users in ChatGPT and Codex

GPT-5.5 Thinking unlocks faster help for harder problems with smarter, more concise answers

Outperforms Claude Opus 4.7 and Gemini 3.1 Pro on most coding and professional benchmarks

More than 85% of OpenAI now uses Codex every week across all company functions

Multimodal Mixture-of-Experts (MoE) architecture with 1T+ total parameters (42B active)

Extended context window up to 1M tokens

Native multimodal perception supporting text, images, video, and audio

Advanced autonomous agent capabilities handling 1000+ tool calls

40-60% better token efficiency compared to Claude Opus and GPT-5.x

ClawEval benchmark: 64% Pass@3 score

SWE-bench Pro: 57.2% task resolution rate

Surpasses Claude 4.6 Sonnet in coding tasks, approaches Claude Opus in agentic performance

Part of the MiMo-V2.5 family alongside MiMo-V2.5 and MiMo-V2.5-TTS

Available via mimo.mi.com with affordable token plans (monthly/annual subscriptions)

27B dense open-source model with Apache 2.0 license

Surpasses Qwen3.5-397B-A17B on all major agentic coding benchmarks

SWE-bench Verified: 77.2 vs 76.2, Terminal-Bench 2.0: 59.3 vs 52.5, SkillsBench: 48.2 vs 30.0

Supports both multimodal thinking and non-thinking modes natively

Native vision-language support for images and video understanding

GPQA Diamond: 87.8, competitive with models several times its size

Compatible with OpenClaw, Claude Code, and Qwen Code coding assistants

Available on Hugging Face, ModelScope, and Alibaba Cloud Model Studio API

Open-source SOTA on HLE w/ tools (54.0), SWE-Bench Pro (58.6), SWE-bench Multilingual (76.7), BrowseComp (83.2), Toolathlon (50.0), Charxiv w/ python (86.7), Math Vision w/ python (93.2)

Long-horizon coding: 4,000+ tool calls, over 12 hours continuous execution

Generalization across languages (Rust, Go, Python) and tasks (frontend, devops, perf optimization)

300 parallel sub-agents x 4,000 steps per run (up from K2.5: 100 / 1,500)

Proactive Agents: powers OpenClaw, Hermes Agent for 24/7 autonomous ops

Claw Groups research preview: bring your own agents, command friends bots & humans in the loop

API PRICING (use these EXACT values): kimi-k2.6 — Input $0.16/M tokens (cache hit), Input $0.95/M tokens (cache miss), Output $4.00/M tokens, Context Window 262,144 tokens

Sources: https://platform.moonshot.ai, https://kimi.com/blog/kimi-k2-6, https://huggingface.co/moonshotai/Kimi-K2.6

Live on kimi.com in chat and agent mode, plus Kimi Code at https://kimi.com/code for production-grade coding

Most capable generally available Anthropic model for complex reasoning and agentic coding

High-resolution image support: 2576px / 3.75MP (up from 1568px / 1.15MP) with 1:1 pixel mapping

New "xhigh" effort level for coding and agentic use cases

Task budgets (beta) — advisory token budget across full agentic loops

128K max output tokens, 1M context window at standard pricing

+12 points on CursorBench coding benchmarks vs Opus 4.6

New tokenizer (up to ~35% more tokens per text, improved performance)

Adaptive thinking only — extended thinking budgets removed

Sampling parameters (temperature, top_p, top_k) removed

Pricing: $5/$25 per MTok input/output, batch $2.50/$12.50 per MTok

#1 on SWE-Bench Pro (58.4%), beating GPT-5.4 and Claude Opus 4.6

Post-training upgrade to GLM-5 — same 744B MoE architecture (40B active)

Trained entirely on Huawei Ascend chips — no NVIDIA hardware

MIT license, compatible with Claude Code and OpenClaw

202K context window, strong on cybersecurity (CyberGym 68.7%)

Faster variant of Claude Opus 4.6 with comparable intelligence

New Capybara tier above Opus — the most powerful Anthropic model

93.9% on SWE-bench Verified, 77.8% on SWE-bench Pro

97.6% on USAMO 2026, 94.5% on GPQA Diamond

1M context window, limited preview for ~50 partner organizations

Google's most capable open models, built from Gemini 3 research

Four sizes: E2B, E4B, 26B MoE (3.8B active), 31B Dense

First Gemma release under Apache 2.0 license

Native multimodal, 140+ languages, up to 256K context

Agent-ready with function calling and structured JSON output

Vision + Code model from Z.ai

Multimodal coding capabilities

API only

1M token context window with always-on chain-of-thought reasoning

78.8% on SWE-bench Verified — competitive with Claude Opus 4.6

2-3x faster output speed than Claude Opus 4.6

Free preview via OpenRouter, successor to Qwen 3.5

Mistral's first audio model — direct competitor to ElevenLabs

Zero-shot voice cloning with multilingual support

Real-time streaming capabilities

Open weights under CC BY-NC 4.0 (non-commercial)

Xiaomi reasoning model with strong math and code performance

309B MoE architecture

Self-evolving agent model — first to participate in its own development

56.22% on SWE-Pro, matching GPT-5.3-Codex

57.0% on Terminal Bench 2, GDPval-AA ELO 1495 (highest open-source)

230B MoE (10B active), 200K context, open weights on HuggingFace

Agent Teams for native multi-agent collaboration

Efficient variant of GPT-5.4 with native computer use

Lower cost while maintaining strong reasoning capabilities

First open-source code agent for Lean 4 formal proof engineering

Generates code AND machine-checkable mathematical proofs

119B MoE with 6.5B active, outperforms Claude Sonnet 4.6 on FLTEval

Apache 2.0 license, 15x cheaper than Claude Opus for formal verification

Unifies instruct, reasoning, coding, and multimodal in a single model

119B MoE with 6.5B active parameters, 256K context window

Replaces Magistral (reasoning), Pixtral (vision), and Devstral (coding)

Apache 2.0 license, configurable reasoning parameter

Beta release with parallel agents architecture

500K context window

Iterative improvement via user feedback

Open MoE model from NVIDIA

120B total parameters with 12B active

Strong enterprise performance

Latest OpenAI flagship with 1M token context window

Available in Standard, Mini, and Nano variants

Supports reasoning effort with 4 effort levels

128K max output tokens

Prompt caching with $0.02-$0.25/M cached read

Google's high-efficiency model optimized for high-volume use cases

1M token context window, 65.5K max output

Supports prompt caching, reasoning effort, and reasoning budget

Native tool calling and vision capabilities

Google's latest flagship model

More than doubles reasoning performance over Gemini 3 Pro

Released in preview via Gemini API, AI Studio, and Vertex AI

Beta release with rapid learning architecture — improves weekly via user feedback

256K context window

4-agent parallel reasoning

Medical document analysis added

Most capable Sonnet yet with full upgrade across coding, computer use, long-context reasoning

1M token context window in beta

200K token context window, 64K max output

Supports prompt caching, reasoning effort, and reasoning budget

Native tool calling and vision capabilities

Agentic AI model with built-in tools for web search and code execution

1M token context window

Qwen3.5-Plus hosted; open weights planned

Frontier MoE model with 80.2% on SWE-Bench Verified

Strong coding and agentic capabilities

230B total parameters, 10B activated per token

Major update to the V3 series with 1M token context

671B MoE focused on code generation and reasoning improvements

Open weights on HuggingFace, MIT license

China's first public AI company frontier model

Targets complex systems engineering and long-horizon agentic tasks

On-device multimodal LLM with full-duplex real-time audio, image, video

Built on Qwen3-8B architecture

Gemini 2.5 Flash level performance at only 9B parameters

Most capable agentic coding model from OpenAI

Available via Codex app, CLI, IDE extensions

Optimized for software engineering workflows

Huge leap for agentic planning with parallel subtask execution

Tool and subagent orchestration capabilities

Terminal-Bench record holder

1M token context window, 32K max output

State-of-the-art agentic AI behaviors

Open-source sparse MoE with 3-way Multi-Token Prediction

100-350 tok/s generation speed

Frontier reasoning at low cost

400B sparse MoE with only 13B active parameters

Built in the US with open weights

One of the largest open-source foundation models

Apache 2.0 license

Top-tier reasoning model with adaptive tool use

Retrieves information and runs code during inference

Rivals leading frontier models

Massive 1T MoE with 32B active parameters

First open-weight model to rank #1 on LMSYS Chatbot Arena

2M token context window, 200+ language support

$0.15/$2.50 per 1M tokens, Modified MIT license

India's multilingual LLM — part of sovereign AI initiative

Supports 10+ Indian languages natively

Korea's answer to open frontier models

102B MoE model with 12B active parameters

Fast frontier-class model rivaling larger models at a fraction of the cost

Default model in the Gemini app

Multimodal model from AI2

Fully open weights, data, and code

Xiaomi large reasoning model

309B MoE architecture

Strong on math and code

Improved reasoning and multimodal capabilities over GPT-5.1

Enhanced mental health protections

128K max output tokens

Available on Plus ($20/month), Pro ($200/month), and API

Expert-level performance on 44 knowledge work tasks

Successor to Devstral Small 1, derived from Mistral Small 3.1

Portable coding agent

Apache 2.0 license

Next-gen coding model with top SWE-Bench score

Modified MIT license (free unless high revenue)

Largest Ministral 3 model with vision

Best-in-class text and vision capabilities

Apache 2.0 license

Powerful and efficient model with vision

Best-in-class text and vision at this size

Apache 2.0 license

Tiny and efficient edge model with vision

Runs on phones, drones, and laptops

Apache 2.0 license

Amazon next-gen foundation model

Available via AWS Bedrock

Announced at re:Invent

Sparse MoE with 41B active parameters

Open weights

Strong reasoning and multilingual capabilities

Open-weights model topping global coding and reasoning leaderboards

Includes GLM-4.7 Flash variant

Cost-effective compared to Western competitors

Fully open-source SOTA coding model

230B params MoE architecture, 10B activated per token

SWE-bench score of 74.0%

92% cheaper than Western alternatives

Exceeds Sonnet 4.5 by 4.3% using 48% fewer tokens at max effort

200K token context, 64K max output

Hybrid reasoning with instant or extended thinking

Multimodal: text, image, and audio support

20% accuracy gain, Excel and financial modeling breakthrough

Fully open model with weights, data, and training code

From AI2 research lab

Large 671B MoE reasoning model

Strong on complex reasoning tasks

Reasoning variant of Gemini 3

Deep chain-of-thought for complex scientific problems

Over 50% improvement over Gemini 2.5 Pro

Most powerful Google model — replaces 2.5 series

1M token context window

Advanced multimodal: text, image, video, audio, code

Family of four models with adaptive reasoning

Faster, more conversational, improved coding

Rolled out to all ChatGPT users

Upgraded Kimi model with thinking and reasoning capabilities

Most capable Amazon model

1M context window

Multimodal capabilities

Teacher for distillation on Bedrock

First major Russian-developed large language model on the global stage

From Yandex

Upgraded MiniMax model with improved reasoning and generation

Open weights

First GLM model with native support for China domestic chips

Cambricon and Moore Threads support

FP8 and Int4 quantization

IBM open enterprise model

Hybrid Mamba-2 Transformer architecture

Apache 2.0 license

Anthropic's fastest model with near-frontier intelligence

200K token context window, 64K max output

21K+ tokens per second for prompts under 32K tokens

Supports reasoning budget and effort control

Most cost-effective in the Claude family: $1/M input

Further iteration on V3 series

Enhanced capabilities across all benchmarks

Open weights

Anthropic's best model for coding tasks

1M token context window (beta feature)

64K max output tokens

Strong agentic behavior and computer-use skills

Optimized for efficient coding and parallel processing

Ultra-efficient MoE from Alibaba

80B total, only 3B active parameters

Strong reasoning with minimal compute

Apache 2.0 license

Massive 1T MoE model with open weights

Highly competitive with frontier models

Major Chinese AI milestone

32B activated parameters

Cost-effective: ~$0.15/M input, $2.50/M output

Strong coding performance across 32+ languages

98% cost reduction compared to Grok 4 Standard

40% increase in token efficiency

Real-time search integration via X

$0.20/M input, $1.50/M output

Adds vision to Magistral Medium

Multimodal frontier reasoning

Closed API only

Adds vision to Magistral Small

Multimodal reasoning model

Apache 2.0 license

Latest in the Hermes series

Advanced function calling and structured output

Built on Llama 3.1

Major upgrade to V3 with improved reasoning and coding

Open weights

Frontier-class multimodal model

Competitive with GPT-4o and Claude 3.5

Strong vision and reasoning capabilities

Vision-language model from Z.ai

106B parameters with strong multimodal understanding

Next-generation flagship with major intelligence leap

400K token context window

Built-in reasoning with 4 effort levels

Multimodal: text, image, and video-based reasoning

Available in Standard, Mini, and Nano variants

OpenAI's first open-weight models since GPT-2

20B and 120B variants

Historic open-source move from OpenAI

Upgrade to Claude 4 with improved coding and instruction following

200K token context window

Extended thinking support

Vision and tool calling capabilities

Newest Anthropic model with improved creative writing

Enhanced nuance and multi-step reasoning

Lightweight variant of GLM-4.5

106B MoE, efficient inference on 8x H20 GPUs

Z.ai flagship open MoE model

355B total parameters

Strong reasoning, coding, and agentic capabilities

Claimed cheaper to run than DeepSeek

xAI's most powerful model at the time

Major reasoning leap

Trained on expanded Colossus cluster

Efficient on-device model designed for mobile

Runs on phones and edge devices

Most powerful OpenAI reasoning model

Extended thinking for frontier problems

Update to Mistral Small 3.1

Improved instruction following and reasoning

Apache 2.0 license

Open-source MoE from RedNote (China Instagram)

142B total, 14B active

Performance on par with frontier models at time of release

Mistral reasoning model with extended thinking

Strong STEM performance

Apache 2.0 license

Latest 2.5 Pro with enhanced coding, reasoning, and agentic capabilities

Chinese AI lab flagship with strong long-context

Lightning attention architecture

High-performance model balancing speed and intelligence

200K context window, 64K max output

Best model for complex agents and coding

Native tool calling and computer use

Available on free tier of Claude.ai

Most powerful Anthropic model at launch

Parallel tool use, long autonomous tasks

200K token context window

Extended thinking support

Vision capabilities for image understanding

Mistral dedicated coding model

Optimized for software engineering and agentic coding tasks

Apache 2.0 license

Hybrid SSM+attention architecture

Six model sizes from 0.5B to 34B

Punches above weight class on benchmarks

Apache 2.0 license

Cost-efficient reasoning with controllable thinking depth

#1 Chatbot Arena for speed

Front-tier model, competitive with GPT-4o

Strong multilingual capabilities

Apache 2.0 license

Excellent multilingual performance (Chinese, English, and more)

0.6B to 235B variants with hybrid thinking

119 languages supported

22B active parameters in MoE architecture

Strong coding performance

Apache 2.0 license

Open 32B and 9B multimodal with reasoning

Competitive on vision tasks

Efficient reasoning model

Best cost-performance for coding and STEM

Full o3 reasoning model — successor to o1

Deep chain-of-thought capabilities

Optimized for coding and instruction following

1M token context window

Available in Standard, Mini, and Nano variants

Nano: $0.10/M input, $0.40/M output

Open-weight natively multimodal models

Scout: 109B, runs on single H100 GPU, 10M token context

Maverick: 400B, requires H100 DGX system

Early fusion for native text, image, and video understanding

#1 on LMArena at launch

Built-in reasoning capabilities

1M token context window

Native code execution and Google Search grounding

Best overall model at launch

Open reasoning model based on Llama

253B MoE architecture

Strong enterprise tasks

Adds vision capabilities to Small 3.0

Multimodal, 128K context

Apache 2.0 license

Cohere's 111B flagship model

Enterprise RAG and agentic tasks

Multilingual capabilities

Runs on 2 GPUs

1B/4B/12B/27B variants

Multimodal (text+vision)

Single GPU capable, 128K context

8B bilingual (English + Chinese) model with deep thinking mode

Surpasses Llama 3.1 8B and Qwen2.5 7B on reasoning/knowledge tasks

128K context, trained on 4T tokens with 75%+ cost savings

Apache 2.0 license

Dedicated reasoning model from Qwen team

Strong mathematical and logical reasoning

Apache 2.0 license

Largest OpenAI model at the time

Focus on EQ, creativity, reduced hallucinations

Hybrid reasoning — toggle instant/extended thinking

Best coding model at launch

200K context window, 64K max output

3.8B dense model outperforming 2x-size models (Phi-3.5-mini, Llama 3.2 3B)

128K context, 22 languages, function calling and tool use

Trained on 5T tokens (synthetic + filtered public data + code)

MIT license — smallest Phi model with strong reasoning

Trained on Colossus supercluster (100K GPUs)

Strong reasoning capabilities

Open-source reasoning model rivaling o1

Pure reinforcement learning approach

Caused global market shockwaves

671B MoE architecture

Refreshed Small with state-of-the-art performance

Apache 2.0 license

Truly open: weights + training data + training code + evaluation all released

7B and 13B sizes — 7B competitive with Llama 3.1 8B, 13B with Gemma 2 9B

Trained on 4T–5T tokens, 9-point MMLU increase over OLMo 1

Apache 2.0 license

671B MoE trained for $5.5M — matches GPT-4o/Claude 3.5 Sonnet

Revolutionized cost efficiency

Open-source on GitHub and HuggingFace

Strong coding and mathematical reasoning

1B/3B/7B/10B sizes

Enhanced multilingual and multimodal

Apache 2.0 license

14B excelling at STEM reasoning

Outperforms much larger models on math

Google's model for the agentic era with native image and audio generation

Outperforms Gemini 1.5 Pro at twice the speed

Native tool use including Google Search and code execution

Foundation for Project Astra and Project Mariner

70B matching Llama 3.1 405B performance

Massive efficiency gain

Enhanced reasoning with more compute for complex tasks

Available in ChatGPT Pro tier

Foundation model family: Micro/Lite/Pro/Premier

Multimodal, optimized for AWS Bedrock

Code-specialized model in 6 sizes: 0.5B, 1.5B, 3B, 7B, 14B, 32B

32B variant matches GPT-4o coding ability — state-of-the-art open code LLM

Trained on 5.5T tokens (source code + text-code grounding + synthetic)

300+ programming languages, 128K context with YaRN extension

Apache 2.0 license

Mistral's large multimodal model

128K context, native image understanding at scale

Open weights

Largest open-source Transformer-based MoE model at release

389B total parameters with 52B active per token

256K context window

Outperforms Llama 3.1 405B on benchmarks

Fast and cost-effective model

200K token context window, 8K max output

Multilingual and vision capabilities

$0.80/M input, $4/M output

Ideal for high-volume tasks like chatbots and moderation

Ranked #6 on LMSYS Chatbot Arena at launch — #1 in China

Surpassed GPT-4o-0513 and Claude 3.5 Sonnet in overall ranking

Top-3 in Chinese, Math, Coding, and Hard Prompts categories

Founded by Kai-Fu Lee, proprietary model

First Llama models with vision capabilities — 11B and 90B multimodal variants

Lightweight 1B and 3B edge models for on-device deployment

128K context window, competitive with Claude 3 Haiku and GPT-4o-mini

Drop-in replacements for Llama 3.1 text models

0.5B to 72B range

SOTA open model for coding and math

18T training tokens

Apache 2.0 license

Updated Mistral Small with improved instruction following

22B parameters, Apache 2.0 license

Built on NeMo architecture with native vision support

128K context, Apache 2.0 license

First 'reasoning' model with chain-of-thought at inference

PhD-level science and math performance

Merged DeepSeek-V2-Chat and DeepSeek-Coder-V2 into a single model

236B MoE with 21B active parameters, 128K context

Strong coding and general capabilities in one model

MIT license, available on HuggingFace

Mamba-Transformer hybrid MoE

94B active, 256K context

Fastest long-context model at release

4B MoE and 3.8B variants optimized for edge devices

Phone-capable AI with 128K context window

Improved multilingual support over Phi-3

Strong reasoning for its size class

Competitive with GPT-4o and Claude 3.5 Sonnet

Available on X platform

Korean web giant Naver's flagship LLM optimized for Korean language and culture

Two sizes: HCX-L (largest) and HCX-S (lighter), built on LLaMA 2 architecture

100K context window with Korean-optimized tokenizer

Strong cross-lingual reasoning in Asian languages — Korean, Japanese, Chinese

State-of-the-art text-to-image model from ex-Stability AI founders

12B rectified flow transformer architecture

FLUX.1 [schnell] open under Apache 2.0, [dev] non-commercial

Surpassed closed-source alternatives in image quality

128K context, competitive with GPT-4o and Llama 3.1 405B

12 languages supported

Open weights

Largest open model — 405B parameters

Matches GPT-4 on many benchmarks

128K context window

Co-built with NVIDIA, runs on a single GPU

12B parameters with 128K context window

Drop-in replacement for Mistral 7B with SOTA performance in its class

Apache 2.0 license, strong multilingual support

Strong reasoning from China's national lab

Competitive on math and coding

9B and 27B sizes

Outperforms models 2x its size

Knowledge distillation from Gemini

Surpassed GPT-4o and Gemini 1.5 Pro at launch

2x faster than Claude 3 Opus at lower cost

First open MoE code model matching GPT-4 Turbo on coding

338 programming languages supported

NVIDIA's open model for synthetic data generation

Permissive enterprise license

Major upgrade, 0.5B to 72B range

Competitive with Llama 3 70B

Apache 2.0 license

128K context, 26 languages

Competitive with Llama 3 8B

Open-source GLM-4 series

Specialized code model, 80+ languages

32K context, fill-in-the-middle support

ByteDance's flagship LLM, most popular AI product in China

Available via Doubao app and Volcano Engine API

Supports 50+ application scenarios including voice, vision, and coding

Open-source Seed 1.5 variants released under permissive license

'Omni' model with native audio/vision/text

2x faster, 50% cheaper than GPT-4 Turbo

Real-time voice conversation capabilities

236B MoE with only 21B active parameters

Multi-head Latent Attention for efficiency

Open weights

480B MoE with 17B active parameters

Enterprise-focused, strong on SQL and coding

Apache 2.0 license

Mini/Small/Medium variants

Phi-3 Mini (3.8B) rivals Mixtral 8x7B

Phone-capable AI

Trained on 15T tokens, 8B and 70B sizes

New open-source SOTA with massive community adoption

Large MoE with strong multilingual and code performance

Open weights

Optimized for RAG and enterprise

128K context, 10 languages

Grounded generation capabilities

First production Mamba-Transformer hybrid

256K context, novel SSM architecture

Open MoE with 36B active parameters

Outperformed Llama 2 70B and Mixtral

Apache 2.0 license

xAI's first open-source model

314B MoE under Apache 2.0

Largest open MoE at time of release

Haiku/Sonnet/Opus family

Opus matched GPT-4 on most benchmarks

200K context window, vision capabilities

First Claude Opus model with advanced reasoning

200K context window

Pioneered extended thinking capabilities

Vision and tool use support

Mistral's first flagship commercial model

32K context, top-tier reasoning

Google's open-source model from Gemini research

2B and 7B sizes, strong for its class

1 million token context window — 10x previous record

MoE architecture, processes entire codebases

Most capable Gemini 1.0 model

Beat GPT-4 on 30/32 benchmarks

Powers Gemini Advanced

Open language model in two sizes: 1.6B and 12B

Trained on 2T tokens (Falcon RefinedWeb, RedPajama, The Pile, CulturaX)

Competitive with Mistral-7B despite smaller footprint

Stability AI Community License

Open code LLM in 3 sizes: 3B, 7B, 15B — trained on 4T+ tokens from The Stack v2

600+ programming languages, fill-in-the-middle capability

16K context with sliding window attention

Trained on permissively licensed code only

Korean startup Upstage's open model using depth up-scaling

Topped HuggingFace Open LLM Leaderboard at release

Apache 2.0 license

Open-source MoE matching GPT-3.5 quality with only 12.9B active params

Game-changer for open-source efficiency

Apache 2.0 license

Google's multimodal model family (Nano/Pro/Ultra)

Natively multimodal from training

Community fine-tuned model on Mistral/Yi

Strong at instruction following

Popular for local AI

Founded by Kai-Fu Lee

Strong bilingual (English/Chinese) model

Competitive with Llama 2 70B

Third gen GLM with function calling, code interpreter, and agent capabilities

Mistral 7B fine-tuned with DPO

Showed distilled alignment can match RLHF quality

Outperformed Llama 2 70B on all benchmarks despite being smaller

Sliding window attention

Apache 2.0 license

Alibaba's multilingual model series

Strong on Chinese and English tasks

Open weights

Evol-Instruct tuned Code Llama

Top open-source coding model of its era

Strong on HumanEval

Specialized Llama 2 for code generation

Supports Python, C++, Java, and more

100K context window

First truly open-weight large model for commercial use

7B/13B/70B sizes with RLHF-tuned chat variants

Founded the modern open LLM ecosystem

200K context window

Constitutional AI approach

Strong coding and analysis capabilities

Second generation GLM, 32K context

42% faster inference

Stronger math and coding

Trained on 3.5T tokens of RefinedWeb

Topped the Open LLM Leaderboard

Apache 2.0 license

Google's next-gen model powering Bard/Gemini

Improved multilingual, reasoning, and coding

Commercially usable open-source model

Trained on 1T tokens

Apache 2.0 license

Open-source code LLM trained on The Stack (1T tokens, 80+ languages)

8K context window

Stability AI's open-source LLM family

3B and 7B sizes, trained on 1.5T tokens

CC-BY-SA license

Fine-tuned LLaMA on ShareGPT conversations

Achieved ~90% of ChatGPT quality

Launched the Chatbot Arena

Anthropic's first public model

Constitutional AI for safety

100K context window

Multimodal (text + vision), passed the bar exam (90th percentile)

Massive leap in reasoning over GPT-3.5

~1.8T parameters (MoE estimated)

Fine-tuned LLaMA on 52K instructions generated by GPT-3.5

Showed cheap instruction tuning works

Leaked weights ignited the open-source LLM revolution

Showed small models can match GPT-3

65B parameters

GPT-3.5 with RLHF in a chat interface

Reached 100M users in 2 months

Defined the AI era

Instruction-tuned T5

Demonstrated instruction tuning dramatically improves task generalization

First 100B+ open-source multilingual model

Built by 1000+ researchers across 70+ countries

46 languages supported

Meta's open-source GPT-3 equivalent

Full model weights released for research

175B parameters

EleutherAI's 20B open model

First glimpse that local LLMs could scale to GPT-3 territory

Predecessor to today open-source ecosystem

540B parameter model

Breakthrough capabilities in reasoning, code, and multilingual tasks

Proved smaller models trained on more data outperform larger undertrained ones

Redefined scaling laws for LLMs

Introduced RLHF for alignment

Pioneered training models to follow human instructions safely

280B parameter model

Extensive analysis of scaling laws across 152 tasks

GPT-3 fine-tuned on code

Powered GitHub Copilot

Proved LLMs could write functional programs

First open model runnable on consumer hardware

6B params, GPT-2 architecture

Widely deployed in early local AI applications

1.6 trillion parameter MoE model

Demonstrated efficient scaling through sparse expert routing

First Mixture of Experts model at massive scale

600B parameters for machine translation

175B parameters — demonstrated few-shot learning without fine-tuning

Sparked the modern LLM revolution

Text-to-Text Transfer Transformer Explained T5: Google's Revolutionary #T5

Unified framework treating all NLP tasks as text generation

Robustly Optimized BERT

Showed BERT was significantly undertrained

Achieved new SOTA with better training

Generalized autoregressive pretraining

Outperformed BERT on 20 NLP tasks

Initially withheld due to misuse concerns — "Too dangerous to release"

Showed emergent text generation quality at scale

Bidirectional Encoder Representations from Transformers

Revolutionized NLP benchmarks

Became the foundation for search engines

First GPT model — decoder-only transformer

Demonstrated generative pre-training for language understanding

Embeddings from Language Models

Contextualized word representations using bidirectional LSTMs

'Attention Is All You Need' paper introduces the Transformer architecture

The foundation of all modern LLMs