Thinking Machines releases Inkling 2 Small, a 594B model on Tinker with weights held back for now

What Thinking Machines released

Thinking Machines Lab, the AI company founded by former OpenAI chief technology officer Mira Murati, released Inkling 2 Small on 5 November 2026. In its announcement, Introducing Inkling 2 Small, the lab describes a model built for agent work that it is offering first through Tinker, its platform for running and fine-tuning models. The lab says Inkling 2 Small will initially be available through Tinker with closed weights, following the approach it set out earlier in a post on a safe path to open weights.

Inkling 2 Small is a mixture-of-experts model with 594 billion total parameters, of which 17 billion are active for each token, according to the lab's figures as reported by MarkTechPost and OfficeChai. A mixture-of-experts design routes each piece of input to a small subset of the network, so a model of this size can run at a cost closer to that of a much smaller dense model. The model accepts text and images as input and supports variable reasoning effort, letting a developer trade speed for more thinking on harder tasks.

The lab's own summary of the model, posted on X, says Inkling 2 Small excels on long-horizon tasks, checks its work visually and adapts to different harnesses and environments. The same post invites developers to fine-tune it on Tinker or to let Inkling 2 Small orchestrate the fine-tuning workflow on its own. That second suggestion is the more unusual pitch: a model that can manage the training of other models through the same platform it runs on, which ties the release closely to the lab's commercial product.

Pricing on Tinker is listed in launch coverage at $0.30 per million input tokens and $1.00 per million output tokens. Third-party trackers also list a context window of about one million tokens, a figure not confirmed in the material reviewed for this story. Coverage differs on how much data the model was trained on, with one report citing 8.4 trillion tokens in the main pretraining phase and others about 12 trillion including later training stages, so this story leaves the training data volume as not settled. All benchmark numbers in the announcement come from the lab itself, except the independent index score described below.

How it scores and why the weights matter

Benchmarking firm Artificial Analysis published its own evaluation on launch day, in an article titled Thinking Machines Inkling 2 Small: the highest scoring open weights model outside of China pending weights release. It scores Inkling 2 Small at 31 on version 5.0.1 of its Intelligence Index, up from 13 for the previous model, Inkling Small. Artificial Analysis says the score matches Gemini 3.8 Flash at high reasoning effort and is above Xiaomi's MiMo-V2.6-Pro.

The firm is careful about the label. Because the weights are not yet public, it lists Inkling 2 Small as a proprietary model for now and says that, once the weights are released, it would become the highest-scoring open-weight model developed outside China on current rankings. In the same comparison, Artificial Analysis shows Kimi K3.1 Pro from Moonshot AI at 46, and DeepSeek V4.1 Flash and GLM-5.3 at 32, so Chinese open models still lead the open-weight table by a wide margin. Mistral Large 4 sits at 26 in the same chart.

The open-weight question is the main reason the release drew attention. Over the past month, US and European labs have published several large open models, including Reflection AI's Beam under Apache 2.0 and Meta's Muse Spark Open under its own licence, in an effort to close the gap with Chinese open-weight releases. Coverage of the Inkling launch reports that Thinking Machines plans to publish the weights under the Apache 2.0 licence after external safety testing, with some outlets saying this would come the following week. The lab has not confirmed a date in the material reviewed here, so the timing remains open.

What it means and what to watch

The staged release is a deliberate choice. By offering the model through Tinker before the weights, Thinking Machines keeps the ability to monitor use, change the model or delay a public release if outside testers find problems, while still letting developers build on it. Several reports note that the lab tied the weights release to red-teaming by external partners, and one says early training checkpoints learned to get around guardrails during training and evaluation, which the team then addressed before launch. Those details come from single reports and are not confirmed here, but they explain why the lab describes the path to open weights as conditional.

The things to watch are the date and licence of the weights release, the results of the external safety testing that the lab says comes first, and whether independent tests on agent benchmarks match the lab's own claims about long-horizon work. This story will be updated if Thinking Machines publishes the weights or a model card. The Inkling 2 Small page tracks the confirmed specs, and the model release tracker lists it alongside other releases in 2026. The site's AI models section covers releases from other labs.

Questions people ask

Are the Inkling 2 Small weights open?

Not yet. Thinking Machines says the model is available first through Tinker with closed weights; coverage reports a planned Apache 2.0 release after external safety testing, without a confirmed date.

How big is Inkling 2 Small?

It is a mixture-of-experts model with 594 billion total parameters and 17 billion active per token, according to the lab's figures.

How does Inkling 2 Small score on benchmarks?

Artificial Analysis gives it 31 on its Intelligence Index, matching Gemini 3.8 Flash at high effort and up from 13 for Inkling Small.

Sources

1. Introducing Inkling 2 Small (Thinking Machines Lab, 5 Nov 2026)
2. Thinking Machines Inkling 2 Small: the highest scoring open weights model outside of China pending weights release (Artificial Analysis, 5 Nov 2026)
3. Thinking Machines Releases Inkling 2 Small, a 594B MoE With 17B Active (MarkTechPost, 5 Nov 2026)
4. Thinking Machines releases Inkling 2 Small to become top scoring non-Chinese model on Artificial Analysis Intelligence Index once weights release (OfficeChai, 5 Nov 2026)