# Thinking Machines Releases Inkling 2 Small's Weights Under Apache 2.0, a Week After Its Closed Launch on Tinker Times of AI Desk · 2026-11-07 · Models [https://timesof.ai/2026/11/thinking-machines-inkling-2-small-tinker](https://timesof.ai/2026/11/thinking-machines-inkling-2-small-tinker) > Updated Nov. 13: Thinking Machines Lab published the weights of Inkling 2 Small on Hugging Face on November 12 under an Apache 2.0 licence, in NVFP4 and MXFP8 checkpoints, meeting the release window it set at launch. The main NVFP4 checkpoint needs at least 500GB of aggregated GPU memory, for example two Nvidia B300s or four H200s. The model is also served by inference providers including Nvidia, Fireworks and Together. Original report: the lab introduced the 594 billion-parameter, 17 billion-active multimodal mixture-of-experts model on November 5 on Tinker with closed weights, promising open weights after external safety testing; its benchmark figures are lab-reported. **Updated November 13, 2026:** Thinking Machines has released Inkling 2 Small's weights under Apache 2.0. The new section directly below covers it; the original report follows unchanged. ## November 13: The weights are out Thinking Machines Lab kept to its timetable. On **November 12** it added a note to its launch post saying Inkling 2 Small's weights "are now available for download on HuggingFace", a week after it put the model on Tinker with closed weights pending external safety testing. Hugging Face lists the repository under an **Apache 2.0** licence, the licence the lab had promised. There are two checkpoints. The main one is in **NVFP4**, the precision the model card says it was pretrained, midtrained and post-trained in, and needs a GPU cluster with at least **500GB** of aggregated memory, such as two Nvidia B300s or four H200s. An **MXFP8** version needs at least 800GB; the lab says it scores only one to two percentage points higher on its evaluations and recommends NVFP4 for all uses. The model runs locally with SGLang, vLLM, TokenSpeed and Unsloth, and is offered through inference providers including Nvidia, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare, the lab says. The model card says the lab assessed chemical, biological, radiological, nuclear and cyber risk through internal evaluations, external testing and refusal-suppressed variants, and found the model "materially below frontier capabilities" on its loss-of-control evaluations. Those are the lab's own conclusions. ### Limits of the November 13 update - We have not run the weights or reproduced any benchmark; capability and safety findings remain Thinking Machines' own. - The model card we read did not name the external testers or publish their results. Thinking Machines is again trying to walk the path from a closed fine-tuning API to downloadable weights — this time at a larger MoE and with an explicit safety gate in between. On **November 5** the lab introduced **Inkling 2 Small**, a multimodal mixture-of-experts model with **594 billion** total parameters and **17 billion** active, image inputs and variable effort. That is a jump from [Inkling-Small, the 276 billion / 12 billion-active open MoE it released in July](https://timesof.ai/2026/07/thinking-machines-inkling-small), and a different shape from the [original 975 billion / 41 billion-active Inkling](https://timesof.ai/2026/07/thinking-machines-inkling-open-weights). The new model is available now on **Tinker** for inference and fine-tuning, with closed weights. Thinking Machines says it intends to make the full weights available under an **Apache 2.0** license next week, after external safety testing and red-teaming, following what it calls a safe path to open weights. Checkpoints will be downloadable when those weights are released, the post says. ## What the lab claims Thinking Machines says Inkling 2 Small "far surpasses" its previous models on the **Artificial Analysis Intelligence Index**, citing a score of **31**. Its own comparison table lists **AA Index v5.0 at 30.6**. Evaluations in the post were run at effort 0.99 ("xhigh") and temperature 1.0, with a **1 million-token** maximum trajectory limit, up from the 256,000-token limit used for Inkling Small. We have not re-run those evaluations. On the training side, the lab says the first pre-training phase covered **8.4 trillion** tokens and ran for 68 hours on NVIDIA **GB300 NVL72** racks, natively in **NVFP4**. A later joint reinforcement-learning run lasted **1,075** steps and more than **17 million** rollouts over **nine days** on **2,048** GPUs, which the lab contrasts with separate reasoning and agentic RL runs on the previous Inkling. It describes a vision encoder of **292 million** parameters based on Perception Encoder, and says the model does not support audio. Safety numbers in the post are the lab's. It says Inkling 2 Small's chemical, biological, radiological and nuclear knowledge and capability scores sit below the open-weight frontier, and that a helpful-only variant did not significantly advance that frontier on the dimensions it measured. We are not reproducing evaluation methods or assistance detail. ## Limits - Intelligence-index and benchmark figures are Thinking Machines' citations of third-party boards or its own tables; we have not independently reproduced them. - The prose claim of an AA Index score of 31 does not match the 30.6 listed in the lab's AA Index v5.0 table. - Apache 2.0 weights are an intention timed to "next week" after safety testing, not a release that has already happened. - CBRN and cyber comparisons are the vendor's; they are not an independent safety board result. - Pricing, rate limits and Tinker capacity terms are not specified in the announcement we read. ## Sources - [Hugging Face: thinkingmachines/Inkling-2-Small model card (accessed November 13, 2026)](https://huggingface.co/thinkingmachines/Inkling-2-Small) - [Thinking Machines Lab: Introducing Inkling 2 Small (November 5, 2026)](https://thinkingmachines.ai/news/inkling-2-small/)