Inkling 2 Small claims to be the strongest Western open-weight model, with 17 billion active parameters and benchmark scores that far exceed its predecessor.
The startup from former OpenAI CTO Mira Murati, Thinking Machines, has unveiled Inkling 2 Small, its second model generation. Inkling models are designed as base models for fine-tuning, which could make them especially attractive for specialized applications.
The mixture-of-experts model contains a total of 594 billion parameters, of which 17 billion are active at any given time. It accepts text and images as input and offers an adjustable thinking depth that lets users choose between speed and thoroughness.
On the Artificial Analysis Intelligence Index, Inkling 2 Small scores 31, according to Thinking Machines. That's 18 points above the predecessor Inkling Small (13) and roughly on par with Google's Gemini 3.8 Flash.
Inkling 2 Small scores 31 on the Artificial Analysis Intelligence Index, an 18-point gain over the predecessor Inkling Small (13) and on par with Gemini 3.8 Flash. | Image: Artificial Analysis
The biggest gains show up in agentic tasks, where the model independently executes multi-step workflows. On AutomationBench-AA the score jumps from 5 to 46 percent, and on FrontierCode 2.0 it rises from 5 to 25 percent. On the knowledge-work benchmark AA-Briefcase, Inkling 2 Small hits 1,527 Elo, 615 points above its predecessor. The hallucination rate drops from 61 to 27 percent, but largely because the model answers fewer questions. The response rate falls from 77 to 53 percent.
Costs rise but stay well below the competition
The better results come partly at the cost of higher token usage. Inkling 2 Small produces roughly 150,000 output tokens per task on average, about four times as many as its predecessor. At $1.18 per task it costs nearly six times as much, according to Artificial Analysis. That's still far cheaper than Gemini 3.8 Flash ($7.77) or Claude Sonnet 5.5 ($10.02), and it lands closer to the price range of low-cost Chinese models.
Inkling 2 Small costs $1.18 per task according to Artificial Analysis, far less than Gemini 3.8 Flash or Claude Sonnet 5.5. In the quadrant chart (below) it sits in the sweet spot of low cost and high intelligence. | Image: Artificial Analysis
Thinking Machines says it trained the model with reinforcement learning on hundreds of thousands of long-horizon coding tasks. An automated evaluation system checked whether the code worked and how the model approached each task. On FrontierCode 2.0, Inkling 2 Small reaches 25.4 percent, on DeepSWE 72.1 percent, and on SWEBench Pro v2 (Hard) 86.3 percent. The new Deepseek-V4.1-Flash still leads on most of those tests, hitting 28.5 percent on FrontierCode and 77 percent on DeepSWE.
Inkling 2 Small scores 25.4 percent on FrontierCode v2.0 and 72.1 percent on DeepSWE v1.1. DeepSeek V4.1 Flash leads on individual tests; on SWEBench Pro v2 (Hard), Kimi K3 narrowly tops the field. The newer Kimi K3.1 Pro is absent from the comparison. | Image: Thinking Machines
Vision and design scores lead among open-weight models
Thinking Machines says the model is strongest in text recognition, infographic interpretation, and visual reasoning. On the document benchmark GDP.pdf, which tests whether models can answer questions from complex PDF documents, it scores 25.2 percent, the highest among the compared models. On MMMU Pro, a benchmark for multimodal understanding of academic tasks involving diagrams and figures, Inkling 2 Small reaches 81.6 percent.
On PerceptionBench, which tests the visual perception ability of multimodal models, it leads with 64.7 percent. For visual creation, the model earns a 1,340 Elo rating on DesignArena Website, where users rate AI-generated websites in head-to-head comparisons. That's the highest score among open-weight models. Thinking Machines says the model has also made progress in 3D modeling.
In a somewhat unusual demonstration, the team pitted the model against its own researchers in a blind wine tasting. Working from a photo alone, Inkling 2 Small correctly identified three wines in 19 out of 20 attempts.
Compact training run, open weights coming soon
The total training run is remarkably compact compared with the competition. Across all phases the model processed about 12 trillion tokens, according to Thinking Machines, far fewer than comparable rivals. The reinforcement-learning phase after base training needed only 3 trillion tokens on 2,048 GPUs over nine days.
Thinking Machines says the model sometimes communicates inefficiently, and that's likely a bigger issue, it doesn't always pick the best tools for large codebases. It doesn't support audio and is less optimized for chat. It can also overreact to health-related topics and produce cluttered output.
For now, Thinking Machines offers the model through its Tinker platform. The company plans to release the weights under an Apache 2.0 license once external safety testing wraps up. That's one of the most permissive open-source licenses and an advantage over more restrictive open-weight licenses such as those from Meta or Mistral. Thinking Machines says the model belongs to a family of different sizes, and the company plans to release more variants soon.
AI News Without the Hype – Curated by Humans
Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.
Subscribe now
原文版权归 The Decoder 及原作者所有，本站转载仅供阅读。如需删除，请联系 support@aisa.one。
Studio