The GPT OSS open-weight release<\/a> came earlier. OpenAI released GPT OSS 120B and GPT OSS 20B in August 2025, with listed sizes of 116.8B and 20.9B parameters. These releases establish the context for Inkling 2 Small, where the focus is agentic work, coding, and visual understanding, while the earlier events emphasize model availability and a cybersecurity exploit experiment.<\/p>\n Higher scores come with greater token use and a higher cost per task. Artificial Analysis says Inkling 2 Small produces about 150,000 output tokens per task on average, roughly four times the output of Inkling Small. At $1.18 per task, it costs nearly six times as much as its predecessor.<\/p>\n That cost remains below the alternatives tested by Artificial Analysis. Gemini 3.8 Flash costs $7.77 per task, while Claude Sonnet 5.5 costs $10.02. Inkling 2 Small therefore falls closer to the price range of lower-cost Chinese models.<\/p>\n The lower hallucination rate also requires context. Inkling 2 Small’s rate falls from 61 percent to 27 percent, but its response rate falls from 77 percent to 53 percent. The model reaches the new rate by declining more questions.<\/p>\n Thinking Machines says reinforcement learning on hundreds of thousands of long-horizon coding tasks shaped the new model. An automated evaluation system checked whether generated code worked and how the model approached each task.<\/p>\n On FrontierCode 2.0, Inkling 2 Small reaches 25.4 percent. It scores 72.1 percent on DeepSWE and 86.3 percent on SWEBench Pro v2 (Hard). DeepSeek-V4.1-Flash remains ahead on FrontierCode at 28.5 percent and DeepSWE at 77 percent, while Kimi K3 narrowly leads the SWEBench Pro v2 comparison. Kimi K3.1 Pro was not included.<\/p>\n Thinking Machines reports leading results among the compared open-weight models for text recognition, infographic interpretation and visual reasoning. It scores 25.2 percent on the GDP.pdf document benchmark, 81.6 percent on MMMU Pro and 64.7 percent on PerceptionBench. It also earns a 1,340 Elo rating on DesignArena Website, the highest score among open-weight models in that comparison.<\/p>\n Inkling 2 Small’s agentic and visual results may matter when a website depends on a model that can work through a multi-step process, interpret diagrams, read complex PDFs or assess generated web designs. Those capabilities could support analysis of a site’s content, design assets or technical pages, although the available information does not establish specific improvements to search rankings.<\/p>\n For a site audit, token volume, cost and response behavior are equally important to consider. A model that generates about 150,000 output tokens per task may be expensive for repeated scans, while a 53 percent response rate means some requests may finish without an answer. A practical evaluation should therefore track completion rates, total task cost and the accuracy of visual and document analysis, rather than relying on a general intelligence score alone.<\/p>\n The model’s strong DesignArena Website rating also gives website owners a useful basis for testing. Because Inkling models are designed for fine-tuning, teams could adapt the model to a particular site or evaluation process. Any such use would still need to account for its tendency to select tools inefficiently, produce cluttered output and perform less effectively in chat.<\/p>\n Thinking Machines says the complete training run processed about 12 trillion tokens. The reinforcement-learning phase after base training used 3 trillion tokens on 2,048 GPUs over nine days.<\/p>\n The model is initially available through the proprietary Tinker platform. Thinking Machines plans to release the weights under an Apache 2.0 license after external safety testing is complete. It also says more models in the family will be released soon.<\/p>\n Inkling 2 Small is a 594-billion-parameter mixture-of-experts model from Thinking Machines with 17 billion parameters active at any given time. It accepts text and images and is designed as a base model for fine-tuning.<\/p>\n Inkling 2 Small scores 31 on the Artificial Analysis Intelligence Index, 18 points above Inkling Small’s score of 13. It also improves agentic benchmark results, although it uses about four times as many output tokens and costs nearly six times as much per task.<\/p>\n Thinking Machines plans to release the weights under an Apache 2.0 license once external safety testing is complete. Until then, the model is available through the proprietary Tinker platform.<\/p>\n SEOScanPro runs a full technical audit of a site and shows the measured result behind every check. Open the rank tracker<\/a>.<\/p>\nWhat Does the New Intelligence Cost?<\/h2>\n
How Does It Perform in Coding and Visual Tasks?<\/h2>\n
What Does This Mean for Website Audits?<\/h2>\n
What Are the Training and Availability Details?<\/h2>\n
What to Do Now<\/h2>\n
\n
FAQ<\/h2>\n
What Is Inkling 2 Small?<\/h3>\n
How Does Inkling 2 Small Compare With Inkling Small?<\/h3>\n
When Will Inkling 2 Small Weights Be Released?<\/h3>\n
Try the rank tracker<\/h2>\n
![\"SEOScanPro,](\"https:\/\/seoscanpro.ai\/shots\/og-home.jpg\")
<\/a><\/p>\n