Inkling 2 Small strengthens Western open models
Thinking Machines launches Inkling 2 Small, an open-weight model with 594 billion parameters, including 17 billion active. It makes significant progress on agentic tasks, but its inference cost also rises.
Thinking Machines, the startup founded by former OpenAI CTO Mira Murati, has unveiled Inkling 2 Small, a model with 594 billion parameters of which 17 billion are active per query. According to Artificial Analysis, it leads among Western open-weight models, with a clear leap in agentic and multimodal tasks. The downside is more prosaic: it consumes more tokens and costs more to run, although it remains far from the prices of the most expensive premium models. The weights are to be released later under an Apache 2.0 license, after external security testing.
A model gaining ambition, not just size
Inkling 2 Small belongs to the mixture-of-experts family of models, an architecture that activates only part of the network per response to maintain a balance between capacity and cost. The model accepts text and images, and Thinking Machines highlights an adjustable depth of reasoning, intended to let the user choose between speed and rigor. In practice, the promise is clear: to provide a stronger foundation for fine-tuning, thus for specialized uses in business or research.
On the Artificial Analysis Intelligence Index, the score rises to 31, 18 points higher than the previous Inkling Small. The model also approaches Gemini 3.8 Flash on this same index. The most visible gains appear in tasks where the AI must chain actions, correct its course, or handle complex documents. This is where much of the market for so-called “agent” models is played, and it is no small detail.
Des gains nets sur les usages agentiques et visuels
According to Thinking Machines, the most marked progress concerns the automation of multi-step workflows, coding, and image understanding. On AutomationBench-AA, the score jumps from 5 to 46%. On FrontierCode 2.0, it climbs from 5 to 25%. The same logic applies on AA-Briefcase, a knowledge work-oriented benchmark, where Inkling 2 Small reaches 1,527 Elo, 615 points better than its predecessor.
The model also stands out on perception and document reading tasks. It scores 25.2% on GDP.pdf, 81.6% on MMMU Pro, and 64.7% on PerceptionBench, always according to figures published by Thinking Machines and cited by Artificial Analysis. On DesignArena Website, it achieves 1,340 Elo for visual creation, placing it at the top among compared open models. A small in-house demonstration even tested its wine capabilities, without making it overtly snobbish: from a single photo alone, the model correctly identified three wines in 19 out of 20 attempts.
Stronger, but also more verbose in use
The trade-off comes quickly: Inkling 2 Small generates on average about 150,000 output tokens per task, four times more than its predecessor, and its cost rises to $1.18 per task according to Artificial Analysis. This remains lower than Gemini 3.8 Flash, estimated at $7.77, or Claude Sonnet 5.5, at $10.02, but the bill is no longer negligible. In other words, the model gains measured intelligence, but it has not yet learned to be concise. A rare quality, thus immediately suspect in meetings.
Thinking Machines explains it trained the model with reinforcement learning on several hundred thousand long-horizon coding tasks. The reinforcement phase reportedly used 3 trillion tokens on 2,048 GPUs over nine days, for a total of about 12 trillion tokens over the entire cycle. The startup also says the model remains less comfortable in pure chat, does not support audio, and can produce cluttered outputs or overreact on health topics.
Une sortie encore contrôlée, avec les poids attendus plus tard
For now, Inkling 2 Small is accessible via Thinking Machines’ Tinker platform, in proprietary form. The weights are expected later under an Apache 2.0 license, once security checks are completed. This sequence matters because it places the model in an intermediate zone: usable now, but truly open only after external validation. In a market where many announce openness before delivering it, the nuance is worth noting.
Sources agree on one point: Thinking Machines wants to establish Inkling in a family of models offered in several sizes, with other variants to come. If the “Small” version lives up to its promises, it could become a serious base for teams seeking an open Western model, performing on real work tasks and flexible enough to be adapted locally. The issue thus goes beyond the simple technical sheet. It touches on companies’ dependence on large closed models, and the market’s ability to have inspectable, modifiable, and deployable alternatives without relying on a single foreign API.
Key points
- 594 billion parameters, including 17 billion active.
- Score of 31 on the Artificial Analysis Intelligence Index.
- 25.2% on GDP.pdf, best comparative result.
- 1,340 Elo on DesignArena Website.
- Apache 2.0 planned after external red-teaming.
- Mira Murati leads Thinking Machines.
By the numbers
- 594 billion — total model parameters, Thinking Machines, November 2026.
- 17 billion — active parameters per query, Thinking Machines, November 2026.
- 31 — score on the Artificial Analysis Intelligence Index, Artificial Analysis, November 2026.
- $1.18 — estimated cost per task, nearly six times higher than the previous model, Artificial Analysis, November 2026.
- 12 trillion — tokens processed during total training, Thinking Machines, November 2026.