Monolith-1.0: The New Frontier of Open-Weights Reasoning Models

Introduction: A Paradigm Shift in Open AI

On July 17, 2026, the landscape of artificial intelligence shifted fundamentally. Basalt Labs officially released Monolith-1.0, a model that doesn't just compete with closed-source frontier models but arguably surpasses them in pure reasoning capability. For developers and AI engineers, this represents a historic milestone: the arrival of a state-of-the-art, reasoning-centric model that is fully open-weights under the MIT license.

• Release Date: July 17, 2026
• Provider: Basalt Labs
• License: MIT (Open-Weights)
• Primary Focus: Advanced Logical Reasoning and Mathematical Problem Solving

Architecture: Scaling Efficiency via Sparse MoE

Monolith-1.0 is a masterpiece of architectural engineering. It utilizes a sparsely-activated Mixture-of-Experts (MoE) decoder-only transformer architecture. While the model boasts a massive 1.57T total parameters, it maintains extreme efficiency by only activating 49.5B parameters per token. This 32x sparsity allows for high-level intelligence without the prohibitive inference costs typically associated with trillion-parameter models.

• Total Parameters: 1.57T
• Active Parameters per Token: 49.5B
• Architecture: 128 routed experts + 1 shared expert
• Layers: 80 layers
• d_model: 8,192
• Attention: Grouped-Query Attention (64 query heads / 8 KV groups)

Training and Context: The 1M Token Horizon

The scale of training for Monolith-1.0 is unprecedented for an open-weights release. It was trained on a massive 60 trillion token multilingual mixture, with a heavy emphasis on Chinese-English datasets. The compute budget reached approximately 1.8 × 10^25 FLOPs, ensuring a deep understanding of complex linguistic and logical structures.

To handle massive document sets and long-form reasoning, the model features a 1,048,576-token (1M) context window. This was achieved through a sophisticated two-stage YaRN curriculum, extending from an initial 4K pretraining length to the full 1M capacity, allowing for seamless RAG and long-context analysis.

• Training Data: 60 Trillion Tokens
• Compute: 1.8 × 10^25 FLOPs
• Context Window: 1,048,576 tokens (1M)
• Curriculum: Two-stage YaRN extension

Performance: Shattering Benchmarks

Monolith-1.0 has set a new gold standard for reasoning. As of April 2026, it leads the four most critical public reasoning benchmarks, demonstrating a level of cognitive accuracy that was previously thought to be the exclusive domain of proprietary models. Its performance on 'Humanity's Last Exam' and 'AIME 2025' is particularly noteworthy, signaling a leap in mathematical and scientific reasoning.