Thinking Machines Lab — the startup Mira Murati founded after leaving OpenAI in September 2024 — shipped its first model on July 15, 2026. It's called Inkling, and per [Decrypt's review](https://decrypt.co/373884/review-inkling-mira-murati-first-open-source-ai), it's a mixture-of-experts model with 975 billion total parameters (41 billion active at inference), pretrained from scratch on 45 trillion tokens. Weights are on Hugging Face under Apache 2.0. No regional restrictions on the license.
That's the release. Here's the part that matters if you're wiring a model into an onchain agent: Decrypt reports an MCP Atlas score of 74.1% — a benchmark for autonomous task completion over the Model Context Protocol, the same protocol most agent tool-calling runs through. That's the number that tells you whether a model can actually drive tools, not just answer questions about them.
The numbers, as reported
- ▸MCP Atlas: 74.1% — autonomous task completion via MCP
- ▸SWE-Bench Verified: 77.6% — autonomous GitHub bug-fixing
- ▸Price: $1 / million input tokens, $4.05 / million output tokens
- ▸Available on OpenRouter, and per the review, drops into Hermes and OpenClaw setups without extra configuration
Decrypt frames Inkling as the strongest open-weights model trained from scratch by a Western lab, and says it clears Nvidia's Nemotron 3 Ultra on agentic tasks by a wide margin (74.1% vs. roughly 45% on MCP Atlas). Those are the outlet's benchmark reads, not numbers SatoHub has independently reproduced — treat them as reported, not verified, until someone reruns the eval.
What the review also found
This is the part that should slow down anyone about to swap a model into production. The same review found a 27-billion-parameter model, Bonsai 27B, outperforming Inkling on practical coding tasks — a much smaller model doing better on the thing most agent workflows actually need. Decrypt also flags heavy censorship behavior, including restricted responses on harm-reduction scenarios a production agent might legitimately need to reason through.
So the headline ("best open-weights model in the West") and the practical read ("a much smaller model beats it at coding, and it says no more than you'd like") are both true at once. Scale bought Inkling a strong MCP Atlas score. It didn't buy it the best coding model, and it didn't buy it a permissive one.
Why the MCP number is the one to watch
Most model reviews lead with chat benchmarks or raw coding scores. For an agent that has to call tools, read a wallet balance, hit an API, or execute a multi-step task, MCP Atlas is closer to the real job description. A model can write elegant code in isolation and still fumble the tool-call loop that makes it useful inside an agent. Chinese labs like Qwen reportedly still lead on coding-per-dollar — a reminder that "best open-weights" is never one number, it's a tradeoff you pick per workload.
If you're picking a model for an onchain agent — one that needs to call a DEX skill, check a balance, or execute a payment through an MCP-wired tool — the agentic tool-use score and the price-per-token line matter more than a leaderboard headline. Apache 2.0 and OpenRouter availability mean you can actually test the claim yourself against your own tool-call loop before committing.
What to watch
Watch for independent MCP Atlas reruns outside Thinking Machines Lab and Decrypt's own testing — a benchmark reported by one outlet is a data point, not a consensus. Watch whether the censorship behavior gets tuned in a follow-up release, since that's the kind of thing that quietly breaks an agent mid-task. And if you're comparing models for an onchain agent stack, run the comparison against the frameworks you'd actually deploy on — SatoHub's directory maps the harnesses (OpenClaw, Hermes, and others) a model like this plugs into, so the price-per-task math is something you can check, not take on faith.
Sources
- ▸[Decrypt: "Review: Inkling, Mira Murati's First Open-Source AI"](https://decrypt.co/373884/review-inkling-mira-murati-first-open-source-ai)