Weights released a week after the announcement
Thinking Machines Lab said on Nov. 12 (5:50:38 a.m. Nov. 13 Japan time) through its official account on X that the weights (the trained parameters) of its AI model "Inkling 2 Small" are now available for download on Hugging Face. The license is Apache 2.0.
Inkling 2 Small’s weights are now available for download on HuggingFace under an Apache 2.0 license. Make Inkling your own! https://t.co/KS5zd97XSd
When the company announced the model on Nov. 5 (Nov. 6 Japan time), it first offered it through its model development service "Tinker" with the weights withheld, and said it intended to release the full weights under Apache 2.0 the following week, after completing external safety testing and red-teaming, in which testers play attackers to look for weaknesses. The announcement page now carries an update dated Nov. 12 saying the weights are available on Hugging Face.
In its model card, the company said it released the weights so they can be used for research, fine-tuning and integration into other companies' products. The license field of the model card also lists Apache 2.0.
594 billion total parameters, 500GB or more of VRAM even in the standard format
According to the model card, Inkling 2 Small has 594 billion total parameters, of which 17 billion are used in each pass. It uses a "mixture of experts (MoE)" design that selects 8 of 384 experts, and has 48 layers. It takes text and images and answers in text, with a context window of up to 1 million tokens.
The weights are distributed in two formats. The main NVFP4 format needs a GPU setup with a total of 500GB or more of VRAM (GPU memory), and the company gives two NVIDIA B300 GPUs or four H200 GPUs as examples. The other format, MXFP8, needs 800GB or more, with four B300s or eight H200s as examples. The company said that because pretraining, midtraining and post-training were all done in NVFP4, it is the model's native precision, and it recommends NVFP4 for all uses. It said MXFP8 scores 1 to 2 points higher on evaluations, but that the difference is not meaningful in practice.
Seven companies including NVIDIA offer APIs; OpenRouter lists $0.30 for input
According to the update on the announcement page and the Hugging Face page, the model became available through APIs from NVIDIA, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare along with the weight release. SGLang, vLLM, TokenSpeed and Unsloth support running it on one's own servers, and Miles, SkyRL and Snowflake Cortex Training can be used for fine-tuning. The company's materials do not list the API prices of these providers.
The model page of OpenRouter, a service that routes requests to AI models, lists Inkling 2 Small at $0.30 for input and $1 for output per 1 million tokens. The same page lists the context window as 1,048,576 tokens. OpenRouter is not among the seven companies named by Thinking Machines.
According to the model card, the model can also be fine-tuned and used through an API on the company's Tinker, and the trial interface "Tinker Playground" allows inference only.
