Thinking Machines' Inkling 2 Small: A 594B Open-Weight Model Nearing the Frontier

Key takeaway: Thinking Machines Lab released Inkling 2 Small on November 5, 2026 — a 594-billion-parameter open-weight AI model that scores 31 on the Artificial Analysis Intelligence Index, putting it close to frontier models. It runs today on the company’s Tinker platform, and the full weights are promised under an Apache 2.0 license next week, after external safety testing.

What Thinking Machines just released

Thinking Machines Lab, the AI lab founded in February 2025 by former OpenAI chief technology officer Mira Murati, introduced Inkling 2 Small on November 5, opening the second generation of its Inkling model family. The first Inkling models arrived in July 2026 as open weights; this one starts closed and opens later.

The model is a mixture-of-experts design: 594 billion total parameters, with 17 billion active for each token as it is routed to 8 of 384 experts. It accepts text and image input, and its model card lists a context window of up to 1 million tokens, though Tinker serves it today in 64K and 256K variants.

On benchmarks, the company says Inkling 2 Small scores 31 on the Artificial Analysis Intelligence Index — its highest result yet, and enough that, once the weights ship, it would be the top open model not from China, ahead of Mistral Large 4 “Le Chonk” at 26. It also scored 1,340 on DesignArena’s website board, behind only Muse Wonder and Muse Spark 1.3.

How much does Inkling 2 Small cost to use?

On Tinker, the company’s fine-tuning and inference platform, Inkling 2 Small costs $0.30 per million input tokens, $0.03 for cached input, and $1.00 per million output tokens. Training runs $1.50 per million tokens at 64K context. The company says sampling is about 79% cheaper than the original Inkling, which matters for developers running agents that make many calls in a row.

For now, the model is only available through Tinker, with closed weights. Thinking Machines says it will release the full weights under an Apache 2.0 license next week, after external safety testing and red-teaming. Until the weights land, teams can fine-tune and run the model on Tinker, but they cannot download and host it themselves.

How does Inkling 2 Small compare with other open models?

The comparison that matters most is against other open-weight models Western teams can actually run. Inkling 2 Small’s Intelligence Index score of 31 leads Mistral Large 4’s 26, and the company positions it as the strongest open option outside China once its weights are public. Against closed frontier models it is close but not at the top: its DesignArena score of 1,340 trails Muse Wonder and Muse Spark 1.3. The trade-off is scope — the first-generation Inkling models handled text, images, and audio in one model, while Inkling 2 Small drops audio and gains scale, with total parameters more than doubled from the original Inkling Small’s 276 billion.

Why this release matters

Open-weight models are the practical middle ground between using someone else’s API and building your own model. A company can fine-tune Inkling 2 Small on its own data, run it on its own infrastructure, and keep sensitive work in-house — things that are hard or impossible with a closed API. That is the pitch Thinking Machines has made since its founding: models businesses can customise and operate themselves.

The release also lands in a busy week for the lab. Reports say Thinking Machines is close to closing a $5 billion funding round led by Accel, on top of the $2 billion seed round at a $12 billion valuation it raised in 2025. Those funding reports are unconfirmed, but the direction is clear: the lab is scaling fast, and this model is its bid to be taken seriously as a frontier competitor.

What to watch next

First, whether the weights actually ship next week, and under what license. A delay or a more restrictive license would undercut the open-weights story. Second, independent testing: most figures so far come from the company’s own announcement and test setups it chose, so outside benchmarks after the weights land will show where the model really stands.

Analysis: a strong model, but the open part is still a promise

Analysis — our take, not a verified fact: Inkling 2 Small looks like a genuinely strong model at a fair price. But until the weights are public, the “open” in open-weight is a promise, not a fact. The cautious rollout is reasonable — safety testing before a 594-billion-parameter release is the responsible move — yet teams that need to self-host today should wait for next week’s release rather than build on Tinker alone. If the weights land on schedule and independent tests hold up, this becomes one of the best open models a business can run.

Frequently asked questions

What is Thinking Machines’ Inkling 2 Small?

Inkling 2 Small is a 594-billion-parameter mixture-of-experts AI model released by Thinking Machines Lab on November 5, 2026. It takes text and image input, activates 17 billion parameters per token, and scores 31 on the Artificial Analysis Intelligence Index. It is the second generation of the lab’s Inkling model family.

Is Inkling 2 Small open source?

Not yet. The model launched with closed weights, available only through Thinking Machines’ Tinker platform. The company says it will release the full weights under an Apache 2.0 license next week, after external safety testing and red-teaming are complete. The first-generation Inkling models are already available as open weights under Apache 2.0.

How much does Inkling 2 Small cost?

On Tinker, Inkling 2 Small costs $0.30 per million input tokens, $0.03 for cached input, and $1.00 per million output tokens. Training costs $1.50 per million tokens at 64K context. Once the weights are released, teams will be able to download and run the model on their own hardware at no license cost.

How does Inkling 2 Small compare to other AI models?

Its Artificial Analysis Intelligence Index score of 31 makes it the highest-scoring open model not from China once its weights ship, ahead of Mistral Large 4 at 26. On DesignArena it scored 1,340, behind only Muse Wonder and Muse Spark 1.3.

Who is behind Thinking Machines Lab?

Thinking Machines Lab was founded in February 2025 by Mira Murati, OpenAI’s former chief technology officer. The lab builds open-weight AI models and Tinker, its fine-tuning and inference platform. It raised a $2 billion seed round at a $12 billion valuation in 2025, led by Andreessen Horowitz.

Sources

For more on how AI agents are being used in business, see our guide to AI agents for business automation, and our coverage of Microsoft’s Maia 300 AI chip.