Thinking Machines Lab Releases Inkling 2 Small Multimodal MoE Model

In practical enterprise applications, the model shines on the EnterpriseBench-CoreCraft benchmark from Surge AI. It achieved a remarkable 71.8% score, demonstrating its ability to perform complex, multi-step tasks within a simulated professional environment (specifically a fictional PC parts supplier). This level of agentic performance suggests that Inkling 2 Small is more than a chatbot; it is a functional engine for automated workflows.

The model's lineage is also worth noting. It succeeds the first-generation Inkling-Small (released July 2026), which served as the foundational teacher model for Nous Research's post-training Arete-27B. This lineage of distillation and teacher-student training has clearly paid dividends in the second generation's reasoning density.

• Artificial Analysis Index Score: ~31
• EnterpriseBench-CoreCraft: 71.8%
• Training Data: 17 million agent rollouts
• Comparison: Leads Western open-weight models

API Pricing and Accessibility

To ensure immediate developer adoption, Thinking Machines Lab has partnered with OpenRouter to provide low-latency API access. The pricing model is highly competitive, specifically designed to lower the barrier to entry for startups and researchers looking to build on top of high-parameter models without the massive overhead of self-hosting.

The pricing is transparent and follows a standard per-token structure, making it easy to estimate costs for large-scale RAG (Retrieval-Augmented Generation) or long-context agentic tasks. Given the 1M token context window, developers can process entire codebases or massive document sets with predictable expenditure.

• Input Cost: $0.30 per 1M tokens
• Output Cost: $1.00 per 1M tokens
• Provider: OpenRouter
• Context Window: 1,000,000 tokens

Primary Use Cases

Given its high score in agentic benchmarks and its massive context window, Inkling 2 Small is uniquely positioned for several high-value domains. First, in the realm of software engineering, its reasoning capabilities make it an ideal candidate for autonomous coding agents and complex repository analysis.

Second, for enterprise RAG applications, the 1M token window allows for 'needle-in-a-haystack' retrieval across massive datasets without the need for complex chunking strategies. Finally, its multimodal nature makes it perfect for sophisticated customer service agents that need to 'see' user screenshots or technical diagrams to provide accurate troubleshooting.

• Agentic Workflows & Autonomous Agents
• Complex Code Generation & Refactoring
• Large-Scale Document Reasoning (RAG)
• Multimodal Technical Support

Getting Started

Developers can begin working with Inkling 2 Small immediately. The open weights were released on Hugging Face on November 4, 2026, under the Apache 2.0 license. This includes both the base Inkling-2-Small weights and an MXFP8 quantized variant, which is optimized for deployment on consumer-grade and enterprise hardware with reduced memory footprints.

For those preferring a managed service, the API is live via OpenRouter. We recommend starting with the quantized variant if you are self-hosting to maximize throughput, or utilizing the API for rapid prototyping of agentic loops.

• Weights: Available on Hugging Face (Apache 2.0)
• Quantization: MXFP8 variant available
• API Access: OpenRouter
• Deployment: Local (Hugging Face) or Cloud (API)

API Pricing — Input: $0.30 per million tokens / Output: $1.00 per million tokens / Context: 1M tokens

Sources