Thinking Machines’ new model takes the lead in western open-weight AI

ZDNET’s cardinal takeaways

• Thinking Machines released a competitory caller unfastened model, Inkling 2 Small.
• Most competitory unfastened models travel retired of China.
• The merchandise follows French laboratory Mistral’s Le Chonk unfastened model, which excels astatine coding.

Thinking Machines Lab, a tiny AI startup from ex-OpenAI CTO Mira Murati, has released its latest model, Inkling 2 Small — and according to third-party tester Artificial Analysis, it’s acceptable to springiness Chinese models a tally for their lead. It’s technically proprietary until now, but Thinking Machines said it volition merchandise its weights adjacent week.

Also: Open weights vs. closed: An AI civilian war’s afoot

Chinese companies person led the battalion connected unfastened models implicit the past fewer years. Models from DeepSeek and Moonshot regularly apical leaderboards and rival the capabilities of overmuch much expensive, closed models from OpenAI and Anthropic.

What Inkling 2 Small does

Thinking Machines’ latest exemplary has 594 cardinal parameters and supports representation inputs but outputs lone text. The institution noted its web and crippled plan capabilities and that it evaluated the model’s attack to coding problems alongside the codification it generated during testing.

“The effect is simply a exemplary that is some stronger astatine coding and easier to usage wrong existent cause loops,” the institution said.

Also: Nvidia takes different measurement guardant for US open-weight AI

Inkling 2 Small scored a 31 connected the Artificial Analysis Intelligence Index, a composite trial “for tracking advancement toward artificial wide intelligence” that combines benchmark scores spanning coding, science, and wide categories. A people of 31 is supra mean for models successful its people and is tied with Gemini 3.8 Flash. For context, Gemini 4 Carbon and Claude Opus 5.5 apical the scale astatine 60 and 59, respectively.

The open-weight race

Thinking Machines said it volition merchandise the model’s weights adjacent week, a signifier that is becoming customary. Given the heightened information concerns astir AI successful general, AI companies walk a fewer weeks investigating the to-be-open exemplary with cybersecurity teams and information reviewers earlier releasing the weights, alternatively than starring with the weights retired of the gate.

French AI laboratory Mistral precocious released its Le Chonk exemplary via the aforesaid staggered process — it present keeps pace with OpenAI’s GPT-6.1 Terra successful coding abilities, which is uncommon for western-made unfastened models.

Also: Mistral’s Le Chonk is present a apical unfastened exemplary for coding

“[Thinking Machines’] program to merchandise its weights would marque it the highest-scoring unfastened weights exemplary developed extracurricular China based connected existent rankings,” Artificial Analysis said.

But seldom is simply a exemplary the afloat package.

“Inkling 2 Small is amongst the starring models successful quality and reasonably priced erstwhile comparing to different models of akin price. It’s besides notably dilatory per task and precise verbose,” Artificial Analysis said. The enactment besides called Inkling 2 Small’s pricing “somewhat expensive” astatine $0.30 per 1 cardinal input tokens, since the median outgo for its people is $0.24.

Still, the model’s advancement implicit its predecessor stands retired to Artificial Analysis reviewers: Inkling 2 Small jumped 41% successful automation capabilities and 20% successful coding abilities. However, astir models excel successful definite areas oregon connected circumstantial benchmarks, making it hard to deem 1 oregon different the outright champion coding exemplary overall.

Also: I tried replacing Claude Code with a section AI stack that’s wholly free

Inkling 2 Small besides made immoderate gains successful accuracy, Artificial Analysis said.

“On AA-Omniscience v1.1, the hallucination complaint falls from Inkling Small’s 61% to 27%, portion accuracy rises from 32% to 34% and the effort complaint falls from 77% to 53%,” its reappraisal reported.

Radhika Rajkumar

Senior Editor

Radhika Rajkumar is simply a elder exertion astatine ZDNET based successful New York City. She covers AI, specializing successful safety, privateness and security, policy, education, and synthetic media. She besides leads ZDNET's newsletter strategy. Radhika holds a Masters successful Creative Publishing and Critical Journalism from The New School. See afloat bio