Thinking Machines' Inkling 2 Small Leads Western Open-Weight Models, With Caveats

If you've been waiting for a Western open model that can hold its own against the Chinese releases dominating leaderboards, Thinking Machines has a new candidate. The small AI lab, founded by former OpenAI CTO Mira Murati, has released Inkling 2 Small. Third-party tester Artificial Analysis says it's the strongest open-weight model built outside China on current rankings. That claim comes with a condition, though, and the details matter more than the headline.

Why the open-weight question matters

Right now, the strongest models most people use are closed. You send prompts to a company's servers and pay per token, and you never see the weights, the numerical parameters that define how the model behaves. Open-weight models flip that. Anyone can download them, run them on their own hardware, and tune them for specific jobs.

Chinese labs have set the pace here for a while. Models from DeepSeek and Moonshot regularly top open leaderboards and rival the capabilities of far more expensive closed systems from OpenAI and Anthropic. DeepSeek's R1 drew real attention in January 2025 when it performed similarly to OpenAI's o1, and that moment pushed many investors and enterprise buyers to question how much they should keep paying for closed-model contracts. A strong Western open model changes that math, at least on paper. Companies that want control over their data or want to avoid vendor lock-in now have a stronger option to test.

What Inkling 2 Small actually is

Inkling 2 Small has 594 billion parameters. It accepts image inputs but returns only text. Thinking Machines highlighted its work on web and game design, and said it evaluated the model's approach to coding problems alongside the code it produced during testing. The company's pitch is that the result is "both stronger at coding and easier to use inside real agent loops," meaning workflows where a model calls tools, checks its own output and keeps going without a person stepping in each time.

That agent focus is where I'd look first if you're evaluating it for work. Chat quality is easy to judge by eye. Whether a model reliably completes a multi-step task without wandering off is harder, and it's the thing that tends to separate a demo from a usable tool.

Where it ranks

On the Artificial Analysis Intelligence Index, which combines benchmark results across coding, science and general categories, Inkling 2 Small scores 31. That's above average for its class and ties Gemini 3.8 Flash. The top of the index is a different tier entirely: Gemini 4 Carbon scores 60 and Claude Opus 5.5 scores 59. So this is a strong result for a model of its size, not a frontier-beating one.

The improvement over its predecessor, Inkling Small, is more telling. Artificial Analysis reports a 41% gain in automation capabilities and a 20% gain in coding. Accuracy also moved in the right direction. On AA-Omniscience v1.1, the hallucination rate drops from 61% to 27%, accuracy rises from 32% to 34%, and the attempt rate falls from 77% to 53%. Read that last pair carefully. A lower attempt rate means the model more often declines to answer when it isn't confident, which can look like a weaker result on paper while actually reducing fabricated answers. For anyone who has been burned by a confident wrong answer, that trade is a good one.

Coding is more mixed. Most models excel in some areas and lag in others, and it's hard to name one as the outright best coding model. Inkling 2 Small looks competitive, not dominant.

The catches

Artificial Analysis gave the model credit for being among the leading models on intelligence and reasonably priced against similar models. It also flagged three problems. The model is "notably slow per task," it's "very verbose," and its pricing is "somewhat expensive" at $0.30 per million input tokens, compared with a class median of $0.24.

Those three issues compound each other. A verbose model produces more output tokens, and output is usually the expensive part of the bill. Slow per-task runs also hurt agent workflows, where a model may chain many calls together. If you're building something that runs thousands of tasks a day, a few seconds per task and extra output tokens can add up faster than the sticker price suggests. If you're using it for occasional coding help or research, those costs matter far less.

The staggered release

Thinking Machines hasn't released the weights yet. It says they'll arrive next week, and that the model is technically proprietary until then. Artificial Analysis's "highest-scoring open weights model developed outside China" claim is based on current rankings and depends on that release happening as planned.

This staging is becoming common. Labs often spend a few weeks testing a model with security teams and safety reviewers before the weights go public, since a model that anyone can download is much harder to take back. Mistral followed the same path with its Le Chonk model, which now keeps pace with OpenAI's GPT-6.1 Terra in coding, a rare result for a Western open model. Taken together, Le Chonk and Inkling 2 Small suggest that the gap between Western and Chinese open models is narrowing, at least on coding benchmarks.

What to watch next

My read is that the headline will shift once the weights are public and independent testers can run the model on their own hardware. Benchmarks on hosted endpoints don't always match results on local setups, and community fine-tunes can change the picture quickly. Watch for speed and token-cost numbers from third parties, not just the intelligence score. If the verbosity gets better in later versions, or if the pricing comes down, this becomes a much easier recommendation.

For now, Inkling 2 Small is a serious option for developers who want an open model with a strong coding and agent profile, and who can live with slow, chatty output. If you mainly need speed and low cost per task, the rest of the open field is still worth comparing before you commit.