Disclosure: this article contains no affiliate links. All links are plain links, and we earn nothing if you click them. See our affiliate disclosure.
How this was made: researched and drafted with AI tools from the sources linked in this article, then checked against those sources before publishing.
Quick answer: Thinking Machines has released the weights of Inkling 2 Small for download on Hugging Face under an Apache 2.0 license. The @thinkymachines post went up at 20:50 UTC on November 12, 2026 (04:50 MYT on November 13). The company says the model is a Mixture-of-Experts design with 594B total parameters and 17B active, and its model card says the main checkpoint needs at least 500GB of aggregated GPU memory.
What was released
Thinking Machines added a note to its Inkling 2 Small article, a page dated November 5, 2026: “Nov 12, 2026 Update: Inkling 2 Small’s weights are now available for download on HuggingFace.” Mira Murati wrote in a post on X: “As promised, we’re releasing Inkling 2 Small’s weights, thanks to the hard work of our team and safety testing partners”.

The November 5 article had said: “Inkling 2 Small will initially be available through Tinker, with closed weights.” It added: “We intend to make the full weights available under an Apache 2.0 license next week after the completion of external safety testing and red-teaming.”
There are two checkpoints on Hugging Face, Inkling-2-Small and Inkling-2-Small-MXFP8. Hugging Face’s API listing for the main repo shows the license as apache-2.0 and the repo as not gated. The model card lists the license as Apache 2.0 and calls the model “a general-purpose multimodal model that accepts text and image inputs and generates text outputs.”
Size, context and hardware
Thinking Machines’ article says Inkling 2 Small is “a Mixture-of-Experts model with 594B total parameters, 17B active, and supports image inputs and variable effort.” The model card adds that each token is routed to 8 of 384 experts and that the model supports “a context window of up to 1M tokens.”

On hardware, the model card says: “The main NVFP4 checkpoint requires a GPU cluster with at least 500GB of aggregated VRAM.” Its examples are 2x NVIDIA B300 GPUs or 4x NVIDIA H200 GPUs. The alternative MXFP8 checkpoint needs at least 800GB, for example 4x B300 or 8x H200. The company writes: “We recommend the NVFP4 checkpoint for all use-cases.”
The Hugging Face README lists SGLang, vLLM, TokenSpeed and Unsloth for local inference. The model is also available through Tinker, and the README lists API access through NVIDIA, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare. We did not test any of these. For other open-weight releases we have covered, see Kimi K3.1 Pro Open Weights Live on Hugging Face and Ripple-65B Preview.
Claims and limits
The company says Inkling 2 Small “far surpasses our previous models’ scores on the Artificial Analysis Intelligence Index with a score of 31.” That is Thinking Machines’ own statement; we did not check the index.
The article’s Limitations section says: “this model does not support audio.” It also says the model is “less optimized for the chat experience.” On safety, the model card says the company concluded that Inkling 2 Small “did not present risk of material uplift beyond what’s already available in the open-weight ecosystem.” The model card also lists a Model Acceptable Use Policy, which we did not review.
FAQ
Is Inkling 2 Small open source?
Thinking Machines released the weights on Hugging Face under an Apache 2.0 license, according to its model card and the Hugging Face listing. We did not review the training data or code, and the model card also lists an acceptable use policy.
How big is Inkling 2 Small?
The company says it is a Mixture-of-Experts model with 594B total parameters and 17B active.
What hardware does Inkling 2 Small need?
The model card says the main NVFP4 checkpoint requires at least 500GB of aggregated VRAM, for example 2x NVIDIA B300 or 4x NVIDIA H200 GPUs. The MXFP8 checkpoint needs at least 800GB.
Does Inkling 2 Small handle images and audio?
It accepts text and image inputs and generates text outputs, per the model card. The company says the model does not support audio.
Sources
– Thinking Machines, Introducing Inkling 2 Small (dated November 5, 2026; Nov 12, 2026 update, no time shown)
– Thinking Machines, Inkling 2 Small Model Card
– Hugging Face, thinkingmachines/Inkling-2-Small and MXFP8 repo (API listing and README read November 13, 2026)
– Thinking Machines (@thinkymachines) on X, weights announcement, November 12, 2026, 20:50 UTC
– Mira Murati (@miramurati) on X, post, November 12, 2026, 20:55 UTC
– TechKit, Kimi K3.1 Pro Open Weights; Ripple-65B Preview