The newly released small model is now available via API
Inkling 2 Mini, the AI model that Thinking Machines Lab released on Nov. 20 (Nov. 21 Japan time), was listed on the morning of Nov. 21 Japan time on OpenRouter, a U.S. service that offers many AI models through a single API. According to OpenRouter's API, the model was registered at 11:19 a.m. Nov. 21 Japan time, about six hours after the company announced the release on X at 5:10 a.m. Prime Intellect serves the model, and it is the only provider offering Inkling 2 Mini on OpenRouter.
Inkling 2 Mini is a "looped" model that uses 3.4 billion of its 26.4 billion total parameters per step and runs its middle layers three times. The company has released the weights and says the model is small enough to run on a laptop. ORIENS covered the specifications and benchmarks in its release article.
Some of you thought that Inkling 2 Small was too big at 594B parameters. We heard you: Inkling 2 Mini is a looped transformer with just 26B parameters, small enough to run on your laptop, powerful enough for agentic work with a score of 20.1 on AA Index.
Pricing is $0.10 input, $0.40 output
The price listed in OpenRouter's API is $0.10 per million input tokens and $0.40 per million output tokens. Cache reads, which reuse input that has already been processed, cost $0.02 per million tokens. The context window is 1,048,576 tokens, and the model takes text and images as input and answers in text.
The same Prime Intellect charges $0.15 input and $0.50 output on OpenRouter for Inkling 2 Small, which has 594 billion total parameters. Five other providers also serve Inkling 2 Small: Modal, Nebius, Together and Cloudflare list $0.30 input and $1 output, and DeepInfra lists $0.285 input and $0.95 output.
- Inkling 2 Mini
- Inkling 2 Small
At the time of retrieval, the pricing panel on OpenRouter's model page showed $0.093 input and $0.371 output. The page says it shows the average price customers actually pay next to the prices providers post, and that caching and discounts often make the price actually paid well below the listed one.
One provider, and a shorter output cap than Small
On OpenRouter, Inkling 2 Mini is served only by Prime Intellect. A single response is capped at 65,536 tokens, shorter than the 943,718 tokens the same provider allows for Inkling 2 Small. The listing shows the served weights in 4-bit (fp4) format.
Supported features include tools for calling external tools, reasoning_effort for choosing how deeply the model reasons, and structured_outputs for answering in a fixed format. Thinking Machines says the model is capable of agentic work.
Developer makes no mention of OpenRouter
Thinking Machines' announcement page does not mention OpenRouter or Prime Intellect. Among the company's X posts from the release through shortly after 9 a.m. Nov. 21 Japan time, none of those that could be checked mentioned availability on OpenRouter. An announcement of the listing by OpenRouter itself also could not be confirmed as of Nov. 21.
The access options the company listed at release were its own model development service, Tinker, and software such as SGLang, vLLM and Ollama for running the model on local hardware. At 9:08 a.m. Nov. 21 Japan time, the company said on X that the research platform NDIF is hosting Inkling 2 Mini along with a research model whose number of loops can be changed. A guide for reading and changing the model's internal values across loops is also available, the company said. At 9:14 a.m., it posted that Ollama's software lets users install and run the model locally with a single command.






