THE CRUNCH
Thinking Machines, the startup founded by former OpenAI CTO Mira Murati, has unveiled Inkling 2 Small, its second model generation. The company claims it is the strongest western open-weight model, with 17 billion active parameters drawn from a 594-billion-parameter mixture-of-experts architecture. On the Artificial Analysis Intelligence Index it scores 31, according to Thinking Machines, 18 points above its predecessor and roughly on par with Google's Gemini 3.8 Flash.
The biggest gains are in agentic tasks, where the model independently executes multi-step workflows: its AutomationBench-AA score jumps from 5 to 46 percent, and FrontierCode 2.0 rises from 5 to 25 percent. Visual work is another strength, with Thinking Machines citing top scores among compared models on document, infographic and perception benchmarks.
The improvements come with trade-offs. The hallucination rate falls from 61 to 27 percent, but largely because the model answers fewer questions, with the response rate dropping from 77 to 53 percent. It also uses roughly 150,000 output tokens per task, about four times as many as before, making it nearly six times more expensive at $1.18 per task, according to Artificial Analysis. That is still far cheaper than Gemini 3.8 Flash ($7.77) or Claude Sonnet 5.5 ($10.02).
Thinking Machines says it trained the model with reinforcement learning on hundreds of thousands of long-horizon coding tasks, using a comparatively compact run of about 12 trillion tokens in total. Known weaknesses include inefficient communication, poor tool choices on large codebases, no audio support and less optimisation for chat.
For now the model is available through the company's Tinker platform, with weights planned under a permissive Apache 2.0 licence once external safety testing finishes. More variants in the family are promised soon.