Thinking Machines publishes Inkling 2 Small, a mixture-of-experts model with open weights
Thinking Machines Lab made the open weights for Inkling 2 Small available on November 12, 2026, one week after the model debuted on its Tinker platform with closed weights. The model is a sparse mixture-of-experts design with 594 billion total parameters, of which 17 billion are active for each token. The weights use the Apache 2.0 license, which allows commercial use, modification and redistribution. The lab had said at launch that the release would wait for external safety testing and red-teaming, consistent with a staged-release framework it published in July.
On the Artificial Analysis Intelligence Index the model scores 31, up 18 points from Inkling Small at 13, and the lab says that puts it roughly level with Gemini 3.8 Flash. On DesignArena, a human preference board, it placed fourth overall and first among open-weight models at launch, with an Elo of 1,376. Serving it takes real hardware: the main checkpoint needs at least 500GB of aggregated VRAM, such as two B300 or four H200 GPUs.
The model card lists research, fine-tuning and integration into third-party products as intended uses. Third-party inference providers also serve the model, and Scale AI ran safety evaluations on it. The lab concluded the model poses no material uplift beyond what is already available in the open-weight ecosystem.
Why it matters
For builders, the license is the main change. Apache 2.0 lets a team host the model on its own hardware, fine-tune it and ship it inside a product without a per-token contract, which shifts both unit cost and where sensitive data flows. The trade-off is infrastructure: a 594 billion parameter model is a multi-GPU job, so many teams will reach it through hosted providers rather than their own racks.
For investors and sellers of closed APIs, the release adds pressure at the top of the open-weight tier. Arena rankings measure preference, not reliability on a given workflow, so buyers should run their own evaluations before swapping the model into a pipeline. Published weights cannot be recalled, so a company that depends on the model can pin a version rather than rely on the vendor keeping a hosted endpoint stable.
Key points
- 1Weights went public on Hugging Face on November 12, 2026, under the Apache 2.0 license.
- 2The mixture-of-experts model has 594 billion total parameters and 17 billion active per token.
- 3Artificial Analysis scores the model 31 on its Intelligence Index, up from 13 for Inkling Small.
- 4DesignArena ranks it fourth overall and first among open-weight models, a preference signal only.
- 5The main NVFP4 checkpoint needs at least 500GB of aggregated VRAM, such as two B300 GPUs.
