Thinking Machines Lab released the weights of Inkling 2 Small on Hugging Face on Nov 12, 2026, a week after the model launched with closed weights on Tinker, the company’s fine-tuning platform (Thinking Machines Lab). The license is Apache 2.0 (model card).
- total parameters
- 594B
- active parameters
- 17B
- tokens of context
- 1M
What happened
Thinking Machines introduced Inkling 2 Small on Nov 5 as its second model generation (Thinking Machines Lab). It is a Mixture-of-Experts model, a design that sends each token through a few of many smaller expert networks (here 8 of 384), and it takes text and images in and writes text out (model card). At launch the company said it would publish the full weights the next week, after external safety testing and red-teaming (Thinking Machines Lab).
The Nov 12 update says the weights are now on Hugging Face, where the company lists two checkpoints, Inkling-2-Small and Inkling-2-Small-MXFP8 (Hugging Face). The model also works with Unsloth, SGLang, vLLM and TokenSpeed, can be post-trained with Miles, SkyRL and Snowflake Cortex Training, and is offered by API through NVIDIA, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare (Thinking Machines Lab).
Why it matters
The model card says the open weights are meant for research, fine-tuning and use inside other developers’ products (model card). Running it takes a GPU cluster. The main checkpoint, in NVFP4 (the number format the model was trained in), needs at least 500GB of combined GPU memory, such as 2 NVIDIA B300 or 4 NVIDIA H200 GPUs. The MXFP8 checkpoint needs at least 800GB, and the company says its 1-2 point gain on evals is not meaningful for practical use (model card).
The sources give no download size, and the benchmark scores above come from Thinking Machines’ own testing. I found no independent check of them in these sources.




