Summary: On November 12, 2026, Thinking Machines announced on X that Inkling 2 Small’s weights are downloadable on Hugging Face under an Apache 2.0 license, closing with “Make Inkling your own!” Cloudflare’s docs show the same model running as a hosted option on Workers AI, so you can either pull the weights or call it behind an API. For builders, that combination means fine-tuning freedom on one side and zero-setup inference on the other.
What the release actually covers
The announcement from Thinking Machines is short: weights, Hugging Face, Apache 2.0. Per Cloudflare’s model page, Inkling 2 Small is a multimodal Mixture-of-Experts model from Thinking Machines Lab with 594B total and 18B active parameters, a 1,048,576-token context window, text and image inputs, and built-in reasoning plus tool calling.
What the sources do not establish: benchmarks, quality comparisons against other models, or anything about training data. I’ve covered Thinking Machines’ earlier move into real-time interaction models, and the pattern is the same here — the release arrives first, the eval story arrives later, if at all.
Running it two ways
The hosted path is Cloudflare Workers AI under the model ID @cf/thinkingmachines/inkling-2-small. Pricing is $0.30 per M input tokens, $1.00 per M output tokens, and $0.10 per M cached input tokens. It’s not on Workers Free billing — you need the Paid plan or prepaid AI Gateway credits. Cloudflare lists zero data retention, vision, function calling, and reasoning effort levels from none up through high (the default) and max. There are OpenAI-compatible /v1/chat/completions and /v1/embeddings endpoints, so existing client code mostly plugs in without changes.
A minimal call looks like this:
curl https://api.cloudflare.com/client/v4/accounts/$CLOUDFLARE_ACCOUNT_ID/ai/run/@cf/thinkingmachines/inkling-2-small \
-X POST \
-H "Authorization: Bearer $CLOUDFLARE_AUTH_TOKEN" \
-d '{"messages":[{"role":"user","content":"hello"}]}'If you’d rather not provision anything, the Workers AI LLM Playground lets you test the model in a browser without authentication. Self-hosting the full 594B parameters is a different story — realistically a team with serious GPU infrastructure. The practical next step: preview it in the Playground, then decide whether the Apache 2.0 weights justify the hosting bill.
Sources
- Thinking Machines on X: Inkling 2 Small’s weights are now available for download on HuggingFace under an
- inkling-2-small
View the original post on X