Mira Murati's Thinking Machines Lab just launched its second-generation AI model, Inkling 2 Small — and it is built for one job above all: long, messy coding work inside
AI agents. The model packs 594 billion total parameters, with 17 billion active at a time, and it scored 31 on the Artificial Analysis Intelligence Index, the lab's best result yet. For a startup founded only in February 2025, that is a serious statement of intent.
What Inkling 2 Small Can Actually Do
Inkling 2 Small is a mixture-of-experts model tuned for agentic
coding, vision, knowledge work, and reasoning. It takes image inputs, adjusts how much compute it spends on a task, and — unusually for a model like this — it was trained with reinforcement learning on hundreds of thousands of long-horizon coding tasks across multiple agent harnesses. In practice, that means it can use visual feedback to iteratively build and refine websites, games, and 3D designs. The company says it follows instructions better than the first Inkling, thinks through changes more carefully, and even verifies its own work without being asked. It does not support audio yet. Early versions of the system were tested on real developer tasks under the codename Fledge-Alpha on the OpenCode benchmark.
The Open Weights Are Coming Next Week
For now, Inkling 2 Small is available only through the lab's own Tinker platform, with its weights closed. But Thinking Machines says it will release the full weights under an Apache 2.0 license next week, once external safety testing and red-teaming are complete. That matters: Artificial Analysis notes the model would be the highest-scoring
open-weight model developed outside China once those weights land. The open-weight race has been dominated by Chinese labs and a few big US names — a strong new open entry from a well-funded independent lab changes the conversation.
Why This Launch Matters
Thinking Machines raised $2 billion in July 2025 at a $12 billion valuation, backed by Nvidia, AMD, Cisco, and Jane Street — so this is not a scrappy side project. The lab is betting that the next frontier is not chatbots but AI agents that can run long coding jobs on their own, and Inkling 2 Small is its second swing at exactly that. The company is also unusually honest about the rough edges: the model is not yet optimized for conversation, it can be verbose, its text formatting has quirks, and it sometimes misallocates tool calls in environments with very large toolsets. Those are the kinds of flaws that matter less inside an agent loop — and everything in this release is aimed at agent loops.
The Take
Inkling 2 Small will not replace your chatbot. It is a specialist: a coding agent's engine, built to grind through long tasks and check its own work. If the open weights land next week as promised, expect it to show up fast in coding tools and agent frameworks everywhere. The open-model race just got a new front-runner.
Comments
Post a Comment