Thinking Machines has released the Inkling 2 Small model, planning open weights under the Apache license
Thinking Machines (Mira Murati) has released the MoE model Inkling 2 Small (594B/17B active parameters), available via the Tinker platform. AAII score of 31, comparable to Gemini 3.8 Flash. Public weights under the Apache license will come later.
Thinking Machines, a startup founded by former OpenAI CTO Mira Murati, has released the second generation of the Inkling model — Inkling 2 Small. It is a mixture-of-experts model with 594 billion parameters in total, of which 17 billion are active. The model processes both text and images and allows the depth of reasoning to be adjusted according to the trade-off between speed and thoroughness. According to the company, it achieved a score of 31 on the Artificial Analysis Intelligence Index, 18 points more than its predecessor, Inkling Small, and roughly at the level of Google's Gemini 3.8 Flash model. The biggest improvements are in agentic tasks (AutomationBench-AA rose from 5 to 46 percent) and coding benchmarks (SWEBench Pro v2 Hard at 86.3 percent).
The better results come with higher token consumption and cost. The model generates on average around 150 000 output tokens per task, roughly four times that of its predecessor, and the price reaches $1.18 per task according to Artificial Analysis, nearly six times that of Inkling Small. According to the source, however, this is still significantly cheaper than Gemini 3.8 Flash ($7.77) or Claude Sonnet 5.5 ($10.02). DeepSeek-V4.1-Flash remains in the lead on most coding tests, for example 28.5 percent on FrontierCode 2.0 compared to 25.4 percent for Inkling 2 Small.
The company states that it trained the model using reinforcement learning on hundreds of thousands of long-horizon coding tasks, with the overall training involving roughly 12 trillion tokens, and the RL phase running on 2048 GPUs over nine days. According to the company, the model achieves the best scores among the compared open-weight models on visual tasks (GDP.pdf, PerceptionBench, DesignArena Website). Among the stated limitations, the company mentions inefficient communication by the model, worse tool selection on large codebases, lack of audio support, and overreactions to health-related topics.
For now, the model is available exclusively as a proprietary offering via the Tinker platform. The company plans to release the weights under the Apache 2.0 license after completing external safety testing, and states that the model belongs to a family with multiple sizes, with further variants to follow.
Why it matters
According to documented benchmarks, the model offers performance comparable to mid-tier closed models, but at a substantially lower cost per task than Gemini 3.8 Flash or Claude Sonnet 5.5 — relevant for teams weighing the cost of agentic and coding tasks. The planned release of weights under the Apache 2.0 license would represent one of the least restrictive open-weight models on the market, usable for fine-tuning toward specialized applications, if the release actually happens.
Release card
Inkling 2 Small
Thinking Machines
- Specifications
- 594 billion parameters total, 17 billion active
- Inputs
- text and images as input
- Price
- $1.18 per task (according to Artificial Analysis)
- Availability
- Offered as a proprietary model via the Tinker platform; public weights under the Apache 2.0 license are to be released later, after external safety testing is completed
- Artificial Analysis Intelligence Index 31 The score is 18 points higher than its predecessor, Inkling Small, and corresponds roughly to the level of Google's Gemini 3.8 Flash model.
- AutomationBench-AA 46 % The score on agentic tasks rose from 5% for the predecessor to 46%.
- SWEBench Pro v2 (Hard) 86,3 % The figure shows the model's performance on the hard version of the SWE-bench coding test.
- FrontierCode 2.0 25,4 % On this coding test, according to the source, DeepSeek-V4.1-Flash remains in the lead with 28.5%.
- GDP.pdf 25,2 % This is the highest score among the compared models on the test of reading and understanding complex PDF documents.
- agentic tasks with multi-step workflow completion
- long-horizon coding tasks trained using reinforcement learning
- visual tasks such as reading PDF documents and designing websites
- according to the company, the model sometimes communicates inefficiently and does not choose the best tools for large codebases
- lacks audio support, and the model is less optimized for chat
- the model may overreact to health-related topics and produce cluttered output
The source compares the model with its predecessor, Inkling Small, with Google's Gemini 3.8 Flash in terms of overall score, and on price with Gemini 3.8 Flash and Claude Sonnet 5.5; according to the source, DeepSeek-V4.1-Flash remains in the lead on most coding tests.
The card summarizes information from the article and any dated corrections, with a link to the original source. It is not our assessment of the model. It does not yet have a dedicated editorial profile. Model selection and other announcements →
Two audiences, two different impacts
What this means
For individuals
Developers and teams experimenting with fine-tuning can now try the model via the Tinker platform, but should expect higher token consumption and cost than with its predecessor.
More practical updates →For a business
Companies choosing a model for agentic or coding tasks may consider Inkling 2 Small as a cheaper alternative to Gemini 3.8 Flash or Claude Sonnet 5.5, although for now only as a proprietary service via Tinker, with no guaranteed release date for open weights.
DevelopmentCheck the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.