AivexaNewsSearch
AI news for builders and product teamsChecked every hour

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

Collected Nov 6, 2026

Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, has released a new model called Inkling 2 Small. The model is proprietary for now, but the company said it will publish the weights next week. Third-party evaluator Artificial Analysis, which tracks model capability through composite benchmarks, reported that once the weights are out, Inkling 2 Small would rank as the highest-scoring open-weight model developed outside China. The release follows Mistral's Le Chonk, a French open model that the publication notes is competitive with OpenAI's GPT-6.1 Terra on coding.

The headline specification is 594 billion parameters. The model accepts image inputs but produces only text output, so it can look at a screenshot, diagram or design mockup and respond in prose or code rather than generating images. Thinking Machines highlighted web and game design as particular strengths, and said it evaluated how the model reasons about coding problems alongside the code it actually writes. In the company's framing, that combination produces a model that is both stronger at coding and easier to drop into real agent loops — the iterative cycles where a model calls tools, reads results and decides what to do next.

On the Artificial Analysis Intelligence Index, a composite that aggregates benchmark scores across coding, science and general categories, Inkling 2 Small scored 31. That puts it above the average for models in its class and level with Gemini 3.8 Flash. The index leaders are far ahead: Gemini 4 Carbon sits at 60 and Claude Opus 5.5 at 59. The gap is worth keeping in perspective. A composite score of 31 against 60 reflects a model that is competitive within a tier, not one that challenges frontier closed systems on overall capability.

The generational improvement is the more striking number. Compared with its predecessor, Inkling 2 Small gained 41% in automation capabilities and 20% in coding abilities, according to Artificial Analysis. The same reviewer also measured accuracy on AA-Omniscience v1.1, a test aimed at factual reliability: the hallucination rate dropped from 61% for Inkling Small to 27%, accuracy edged up from 32% to 34%, and the attempt rate fell from 77% to 53%. That last figure deserves attention. A lower attempt rate means the model declines to answer more often rather than guessing, which is a common way to cut hallucinations and can be the right trade for agentic use, where a confident wrong answer is more costly than an abstention.

Artificial Analysis was less complimentary about the operational profile. Its review described Inkling 2 Small as notably slow per task and very verbose, and called the pricing somewhat expensive at $0.30 per million input tokens, against a median of $0.24 for models in its class. Verbatim output and slow task completion compound in agents: longer outputs cost more output tokens and take more wall-clock time, and each step in a multi-step loop inherits both penalties. For developers running high-volume pipelines, the input-token gap of six cents per million is small in isolation but multiplies quickly at scale.

The staggered weight release is becoming standard practice. Rather than publishing weights on launch day, labs spend a period testing the to-be-open model with cybersecurity teams and safety reviewers before making it downloadable. Mistral used the same process for Le Chonk. This suggests western labs are converging on a template: ship the hosted product first, gather third-party evaluation, run external safety review, then release the weights. It gives labs a window to catch misuse vectors before a model can be run locally without guardrails, at the cost of delaying the community fine-tuning and self-hosting that open weights enable.

The broader context is a landscape that Chinese labs have dominated for open models. DeepSeek and Moonshot have regularly topped open leaderboards with models that rival far more expensive closed systems, a shift that began drawing serious attention when DeepSeek's R1 performed comparably to OpenAI's o1 in January 2025. That track record has given enterprises a reason to reconsider expensive contracts with US labs. More competitive western open-weight models change the calculus for buyers who want open licensing but prefer a US or European vendor.

A caveat the source itself raises: few models are the full package, and strength on particular benchmarks does not settle which model is best at coding overall. Inkling 2 Small's profile — strong coding and automation gains, improved factual reliability, but slow, verbose and priced above median — reads as a specialist rather than an all-rounder. This suggests teams should benchmark it on their own agent workloads before standardizing on it.

Why it matters: Developers choosing an open-weight model now have a credible non-Chinese option for coding and agent workflows, arriving with independent benchmark data already published. The trade-offs are visible up front: higher-than-median input pricing, slow per-task latency and verbose output argue for testing against real pipelines rather than switching on the index score alone. For product teams, a competitive western open model may also shift procurement conversations, since self-hosting and price predictability become viable without leaving the US or European vendor ecosystem — an inference from the release, not a stated outcome.

Read at ZDNet · AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Another competitive model developed outside of China could change the open-weight landscape.