What is Inkling 2 Small?

Inkling 2 Small is an open-weight mixture-of-experts language model from Thinking Machines Lab, with 594 billion total parameters and 17 billion active, plus image input and a context window of up to 1 million tokens. Weights are on Hugging Face under Apache 2.0. You can try or fine-tune it on Tinker, or run it locally with tools the company lists.

What Inkling 2 Small is for

Thinking Machines Lab introduced Inkling 2 Small on 5 November 2026 as its second model generation. The official announcement describes a mixture-of-experts model with 594 billion total parameters and 17 billion active, with support for image inputs and variable effort. A 12 November 2026 update on the same page says the weights are available on Hugging Face, and the official Hugging Face model card lists the license as Apache 2.0, confirms 594B total and 17B active parameters, and states a context window of up to 1 million tokens, with text and image input and text output. The company's limitations section says this model does not support audio, even though the broader Inkling product page mentions audio for the family.

The announcement focuses on long-horizon autonomous work, agentic coding, vision, knowledge work, and reasoning. It says the model can use visual feedback to iteratively build and refine websites, games, and 3D designs. Official vendor scores include 31 on the Artificial Analysis Intelligence Index and, in the company's table, 96.5% and 81.7% on ARC-AGI v1 and v2. Those figures are Thinking Machines Lab's reported results, not independent re-runs in this session.

A first path to try or run it

A first path, if you only want to try the model, is the official Tinker Playground linked from the announcement. Opening that path in this session led to a Thinking Machines sign-in page, so this profile does not describe a signed-in playground session. If you want local or self-hosted use, the official 12 November 2026 update and the Hugging Face card point to thinkingmachines/Inkling-2-Small, and list integrations with Unsloth, SGLang, vLLM, and TokenSpeed. A small starting task is a bounded coding or image-and-text request, then a check of whether the model follows the instructions and, for vision, whether it describes what is in the image.

For customization, the company says Inkling 2 Small can be post-trained with Miles, SkyRL, and Snowflake Cortex Training, and fine-tuned or sampled on Tinker. The 12 November update also lists API availability at NVIDIA, Fireworks, Together, Baseten, Nebius, Modal, and Cloudflare. Those hosts' prices should be confirmed on their official pages. This profile does not test those endpoints.

Pricing, access, and limits

Weights are offered at no charge under Apache 2.0 on Hugging Face. Paid official access is through Tinker. The official Tinker pricing page, which states prices as of 19 October 2026, lists the Inkling 2 Small training, sampling, and serverless inference rates quoted in the pricing field above, plus storage at $0.10 per GB per month. Confirm live rates on that page before starting a training run, because the printed as-of date is older than this profile's verification date of 17 November 2026.

The announcement originally said the model would first be available through Tinker with closed weights, then that full weights would follow under an Apache 2.0 license after external safety testing and red-teaming. The 12 November 2026 update and the Hugging Face card confirm that the weights are now available under that license. The limitations section says Inkling 2 Small is not the most succinct communicator, does not always choose the best tools in large skill-heavy codebases, is less optimized for the chat experience, and can overuse slashes or overreact to mental-health and other safety-related queries. Those are the company's own caveats.

What stands out

  • Official 594 billion total parameters and 17 billion active, in a mixture-of-experts design, with image input and a context window of up to 1 million tokens
  • Official open weights on Hugging Face under Apache 2.0, as of the 12 November 2026 update
  • Official support for local inference with Unsloth, SGLang, vLLM, and TokenSpeed, and for fine-tuning on Tinker
  • Official Tinker training, sampling, and serverless inference pricing, listed on the live Tinker pricing page
  • Official vendor-reported gains on long-horizon coding, vision, and reasoning, including an Artificial Analysis Intelligence Index score of 31

Before you choose

  • The official limitations section says the model does not support audio, is less optimized for chat, and may overreact to some safety-related queries
  • Benchmark figures in this profile are Thinking Machines Lab's reported results, not independent measurements
  • Tinker prices are labeled as of 19 October 2026 on the official page; confirm them again before paying for training or inference
  • This is not a hands-on test of Tinker, local serving, or third-party hosts

Plans & pricing

Current pricing on thinkingmachines.ai (opens in a new tab)

Sources & further reading

Based on the official documentation below. Features and access can change; check the provider before making a decision.

  1. Thinking Machines Lab: Introducing Inkling 2 Small (5 November 2026, updated 12 November 2026) (opens in a new tab)
  2. Official Hugging Face model card for thinkingmachines/Inkling-2-Small, checked 17 November 2026 (opens in a new tab)
  3. Official Tinker pricing page, labeled as of 19 October 2026, checked 17 November 2026 (opens in a new tab)
  4. Official Inkling product page, checked 17 November 2026 (opens in a new tab)
Something out of date or incorrect?Report a correction

Was this page useful?

Help us make our coverage more useful. This is feedback on the page, not a rating of the tool.

Share this page