Thinking Machines' Inkling 2 Small takes on China's open-weight lead

Thinking Machines Lab's new model, Inkling 2 Small, is set to become the highest-scoring open-weight AI model developed outside China once its weights ship next week, according to benchmarking firm Artificial Analysis. That single sentence reframes a race that, until now, Chinese labs have largely owned.

The startup, founded by former OpenAI CTO Mira Murati, released the 594-billion-parameter model with a staggered rollout. The model is technically proprietary today. The company said it will publish the weights next week, after the now-customary period of security review.

"[Thinking Machines'] plan to release its weights would make it the highest-scoring open weights model developed outside China based on current rankings," Artificial Analysis said in its review.

Why does one open-weight release matter?

Context is everything here. Over the last few years, most competitive open models have come out of China. Models from DeepSeek and Moonshot regularly top leaderboards and rival the capabilities of much more expensive closed models from OpenAI and Anthropic, as ZDNET notes in its coverage of the release.

That dynamic has real stakes. Open-weight models can be downloaded, inspected, fine-tuned, and run by anyone — researchers, startups, governments. If the strongest freely available models all come from one country, that country's labs set the de facto global baseline for what open AI can do. A western entrant at the top of the rankings changes that equation.

Inkling 2 Small scored 31 on the Artificial Analysis Intelligence Index, a composite benchmark described as "for tracking progress toward artificial general intelligence" that combines scores spanning coding, science, and general categories. A score of 31 sits above average for models in its class and ties with Gemini 3.8 Flash. For comparison, Gemini 4 Carbon and Claude Opus 5.5 top the index at 60 and 59, respectively.

What does Inkling 2 Small actually do?

The model has 594 billion parameters. It accepts image inputs but outputs only text. Thinking Machines highlighted its web and game design capabilities, and said it evaluated the model's approach to coding problems alongside the code it generated during testing.

That testing philosophy produced a model the company is willing to make a direct claim about.

"The result is a model that is both stronger at coding and easier to use inside real agent loops," the company said.

The numbers back up the coding claim, at least relative to its own predecessor. Artificial Analysis reviewers found Inkling 2 Small jumped 41% in automation capabilities and 20% in coding abilities over the prior Inkling Small model. The model also made gains in accuracy.

"On AA-Omniscience v1.1, the hallucination rate falls from Inkling Small's 61% to 27%, while accuracy rises from 32% to 34% and the attempt rate falls from 77% to 53%," the Artificial Analysis review reported.

Halving the hallucination rate is the standout figure. For anyone building agent systems — software that lets models act rather than just chat — a model that invents facts 27% of the time instead of 61% of the time is a materially different tool.

Is it cheap and fast?

Rarely is a model the full package, and Artificial Analysis didn't sugarcoat the trade-offs.

"Inkling 2 Small is amongst the leading models in intelligence and reasonably priced when comparing to other models of similar price. It's also notably slow per task and very verbose," Artificial Analysis said.

The organization also flagged pricing as "somewhat expensive" at $0.30 per one million input tokens, given that the median cost for its class is $0.24. That's a 25% premium over the class median — not dramatic, but enough to matter at scale, where token costs compound across millions of requests.

Speed is the harder sell. "Notably slow per task" is a meaningful drawback for agent workflows, where a model may be called dozens or hundreds of times to complete a single job. Verbosity adds cost on the output side too, since longer responses mean more tokens billed and more latency for users.

There's also a caveat on the coding crown. Most models excel in certain areas or on specific benchmarks, making it hard to deem one or another the outright best coding model overall, as ZDNET's coverage points out. Inkling 2 Small's 20% coding gain is measured against its own predecessor, not against every rival in the field.

How does the staggered release work?

Thinking Machines said it will release the model's weights next week, and that staggered schedule is becoming customary practice. Given heightened security concerns around AI, companies now spend a few weeks testing a to-be-open model with cybersecurity teams and safety reviewers before releasing the weights, rather than leading with the weights out of the gate.

It's a middle path between the closed and fully open camps. The model's capabilities are visible immediately through hosted access, but the downloadable weights — the part that makes a model genuinely open and freely reusable — arrive only after scrutiny.

Where does this leave the western open-weight race?

Thinking Machines isn't alone in pushing back against China's open-weight dominance. French AI lab Mistral recently released its Le Chonk model via the same staggered process. Le Chonk now keeps pace with OpenAI's GPT-6.1 Terra in coding abilities — a rare feat for a western-made open model, and one that earned it recognition as a top open model for coding.

So the pattern is forming. Mistral carries the European flag, Nvidia has taken steps forward for US open-weight AI, and now Thinking Machines — a small startup led by one of the most recognizable names in the industry — is posting numbers that put it at the front of the non-Chinese pack before its weights are even public.

For developers and enterprises that want capable open models without depending on Chinese or American closed labs, the field just got one more serious option. The test arriving next week is whether the released weights deliver what the hosted benchmark scores promise — and whether Inkling 2 Small can hold its ranking once the open-weight leaderboard absorbs it.