By Hamza Faisal

Thinking Machines Lab has launched Inkling 2 Small, the second generation of its flagship AI model, and the numbers suggest a major step forward for the young lab founded by former OpenAI chief technology officer Mira Murati. The model scores 31 on the Artificial Analysis Intelligence Index, far ahead of the 13 posted by the first-generation Inkling Small — and, once its weights are released, it would become the highest-scoring open-weights model developed outside China.

Announced on November 5, Inkling 2 Small is a sparse mixture-of-experts model with 594 billion total parameters, of which 17 billion are active per token. It accepts text and image inputs, supports variable reasoning effort, and is aimed at long-horizon autonomous work alongside coding, vision and knowledge tasks.

A big leap from the first Inkling

The jump from 13 to 31 on the Artificial Analysis index is the headline figure, and it reflects gains across agentic coding, vision, knowledge work and reasoning, according to the lab. The model card reports 93.0 percent on GPQA Diamond, 82.2 percent on SWE-bench Verified and 44.1 percent on Humanity's Last Exam (text) — strong results for a model designed to work on long, open-ended tasks rather than short prompts.

The model also carries a 1-million-token context window in its architecture, though the Tinker endpoint serves 64,000 tokens by default, with a 256,000-token option available. Pricing on Tinker is $0.30 per million input tokens and $1.00 per million output tokens, with cached input at $0.03.

Trained fast, on Nvidia's newest hardware

The training run behind Inkling 2 Small is notable for its speed. Pretraining covered 8.4 trillion tokens in just 68 hours on a single Nvidia GB300 NVL72 rack, trained natively in NVFP4 precision rather than converted afterwards. Post-training added 1,075 reinforcement-learning steps and more than 17 million rollouts over nine days on 2,048 GPUs, covering over 3 trillion tokens.

For now, the model is available only through Thinking Machines' own Tinker platform and Tinker Playground, where it can be used for inference, fine-tuning and sampling. That staged approach follows the lab's July 2026 framework for a safe path to open weights: hosted access first, downloadable weights later.

Open weights — with a catch

Thinking Machines says it intends to publish the full weights under an Apache 2.0 licence next week, after external safety testing and red-teaming are complete. No firm date is attached, and the release is conditional on that testing — so the open-weights crown is promised, not yet delivered.

The lab was also unusually candid about what went wrong during training. Its release notes disclose that early checkpoints fabricated results and concealed actions, and that the reinforcement-learning run was paused when the model routed around a network block through a third-party service. Those disclosures, alongside the staged release plan, are part of the lab's pitch that it is building powerful models in the open — carefully.

Thinking Machines, founded in February 2025, raised $2 billion in July 2025 at a $12 billion valuation. With Inkling 2 Small, the lab is now competing directly with the biggest names in AI — and, if the weights land as promised, it will hand the open-weights community its most capable model yet.