Skip to content

Thinking Machines opens the weights to Inkling 2 Small

Inkling 2 Small, a 594-billion-parameter mixture-of-experts model, is now on Hugging Face under Apache 2.0, a week after launching as a closed, Tinker-only API.

Thinking Machines, the AI lab founded by former OpenAI CTO Mira Murati, released the weights to Inkling 2 Small on 12 November. The model launched a week earlier as a closed, Tinker-only API; this is the open-weight drop the lab had promised once external safety testing finished. It’s on Hugging Face under an Apache 2.0 licence, which permits commercial use and redistribution.

Inkling 2 Small is a mixture-of-experts model with 594 billion total parameters, of which 17 billion activate per token, and a context window of up to 1M tokens. Running the native NVFP4 checkpoint takes around 500GB of aggregated VRAM, for example two Nvidia B300s or four H200s. It already works in vLLM, SGLang, Unsloth and Hugging Face Transformers, and is hosted by providers including NVIDIA, Fireworks, Together and Baseten. Thinking Machines puts its score at 31 on the Artificial Analysis Intelligence Index.

Why it matters: this is a model close to frontier scale, from a lab with serious backing, now available to self-host under a licence with no restriction on commercial use. That’s a real alternative to a metered API for anyone weighing whether to build on a vendor’s hosted model or own the weights outright.

The caveat: 500GB of aggregated VRAM means this is a cluster, not a laptop, so it’s realistically an option for teams with multi-GPU infrastructure or who use one of the hosted providers. The only benchmark numbers available right now are Thinking Machines’ own.

Share

Related reading

More from The Wire

All briefs