Murati's Inkling drops 975B MoE, but your laptop still files for unemployment 😮‍💨

Mira Murati's Thinking Machines Lab released Inkling last week, its first open-source artificial intelligence model, positioning it as the strongest from-scratch Western entry in a field currently led by Alibaba's Qwen, Z.ai's GLM, and Moonshot AI's Kimi. Inkling is published in full on Hugging Face under the Apache 2.0 license, with no regional restrictions. The model uses a mixture-of-experts architecture totaling 975 billion parameters with 41 billion active at inference, was pretrained on 45 trillion tokens, processes text, images, and audio, and supports a 1-million-token context window.

Pricing on OpenRouter is set at $1 per million input tokens and $4.05 per million output tokens, with full compatibility for any Hermes or OpenClaw setup that already routes through OpenRouter. Benchmark numbers back the launch: Inkling scored 74.1% on MCP Atlas, roughly 30 points above Nvidia's Nemotron 3 Ultra, and posted 77.6% on SWE-Bench Verified, against Nemotron's 70.7%. The hardware footprint, however, places Inkling firmly in datacenter territory rather than consumer machines.

Hands-on testing showed uneven results on coding tasks. A 1,955-word prompt to build a zombie shooter produced a blank screen, while a simplified 99-word version delivered a runnable game with correctly mapped keystrokes. Monsters appeared as rectangles and spheres with no background or visible play field, and the most demanding test produced code that did not run. The model interface prominently advertises full privacy, a feature the lab is highlighting for end users.

The wider leaderboard context reinforces why Murati's team framed Inkling as a Western counterweight: Mistral's April release entered a board dominated by Chinese models, and Nvidia's Nemotron remained the only U.S.-trained entry well short of state-of-the-art standing. For raw coding performance per dollar, Chinese open-source models are still reported as ahead, leaving Inkling's clearest demonstrated edge in agentic tool use rather than local development workflows.