Thinking Machines Lab to release 594B parameter open model next week

A new open model contender

Thinking Machines Lab, the AI research company founded by former OpenAI CTO Mira Murati, has announced its second-generation model, Inkling 2 Small. The company plans to release the model weights next week following the completion of external safety testing and red-teaming. Red-teaming is a process of adversarial testing where researchers try to find vulnerabilities or biases in the AI system.

The model is designed to be a high-performing open model from a company based outside of China. While many U.S. companies have historically followed a closed strategy by keeping their model weights private, Thinking Machines Lab intends to open its weights to the public. This move is expected to significantly impact the developer community, where Chinese companies have recently released several high-performance models for free.

The release will occur under the Apache 2.0 license, which is a standard license that permits a wide range of uses. This includes commercial use by various companies and research institutions. By making the weights available, the company allows developers to host and run the model in their own private environments rather than relying on a third-party API.

Architecture and performance gains

Inkling 2 Small is a Mixture-of-Experts (MoE) model, a technique where only a small portion of the model's total parameters are active during any single inference. Although the model has 594 billion total parameters, it only uses 17 billion active parameters at inference. This allows the model to maintain a large knowledge base while remaining more efficient to run than a dense model of the same size.

The company reported that the model earned a 31 on the Artificial Analysis Intelligence Index v5.0.1, marking a major leap from the previous generation's score of 13. These gains are attributed to improvements at every stage of the training process. The developers also integrated large-scale reinforcement learning to boost the model's reasoning, knowledge work, and vision capabilities.

In specific coding benchmarks, the model reached 72.1% on DeepSWE v1.1 and 86.3% on SWEBench Pro v2 (Hard). It also handles image inputs through a 292-million-parameter vision encoder. These results suggest that the model is capable of handling complex tasks that require both logical reasoning and the ability to process visual data.

Benchmarking against rivals

Third-party evaluations have already begun to show how Inkling 2 Small compares to other major models in the industry. It notably beat GPT-6.1 Sol on DesignArena, a metric used to measure the design capabilities of different AI models. Additionally, the model's overall performance reached a level comparable to Gemini 3.8 Flash in various tests.

The company also highlighted that the model has achieved scores comparable to the most advanced models currently coming out of China. This is significant because Chinese models have been rapidly gaining a presence in the open-model landscape. By offering a high-performing alternative from a U.S. company, Thinking Machines Lab aims to provide a different path for developers seeking top-tier performance.

One of the most notable features of the new model is the large headroom for performance gains through additional training. The company provided graphs showing that Inkling 2 Small can raise its performance more cheaply than competing models like DeepSeek-V4.1-Flash-1017. This suggests that the model is built to be highly efficient for developers who want to customize it without spending excessive amounts on compute.

Availability and next steps

For now, the model is being offered in a closed form through the company's development platform, Tinker. Users can currently try the model in the Tinker Playground, where fine-tuning and sampling options are also available. However, the weights will remain private until the safety evaluations and red-teaming are finished next week.

Once the weights are released, the company plans to make checkpoints downloadable for the public. They also expect to offer the model on additional platforms within a few weeks of the official release. This rollout will allow a broader range of users to integrate the model into their own applications and workflows more easily.

Market observers note that the competition for AI models is moving toward better efficiency relative to training costs and the simplicity of fine-tuning. The metrics presented by Thinking Machines Lab suggest that even developers without massive computing resources can improve performance on limited budgets. The final adoption of the model will likely depend on the actual inference costs and the specific terms of the license once they are fully detailed.