Thinking Machines Lab releases first model Inkling, a 975B open-weights MoE

Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has shipped its debut model: Inkling, a 975-billion-parameter mixture-of-experts transformer that activates 41 billion parameters per token. Wired reports it was pretrained on 45 trillion tokens spanning text, images, audio and video, and offers a one-million-token context window — a genuinely frontier-scale, natively multimodal open-weights release from a lab that had been operating in stealth.

The decision to release open weights is the headline strategic choice. It aligns Thinking Machines with the surging open-model camp — Hugging Face's Clement Delangue argued this week that 'the real AI race may no longer be at the frontier' but in open, specialized models, noting a new repo is created on the platform every seven seconds. Murati releasing open weights signals a philosophical break from her former employer's closed approach; Forbes framed her direction as closer to China's open-weight playbook.

MoE at 975B total / 41B active is a bet on capability-per-inference-dollar, competing with DeepSeek's open models, Meta's Llama line, and Alibaba's Qwen. The multimodal pretraining and 1M-token context put it nominally in the same conversation as frontier closed models, though independent benchmarks are the real test. HN buzzed heavily (742 points) over the release.

Caveats: total parameter counts and token budgets don't guarantee quality, and running a 975B MoE requires serious infrastructure even with sparse activation. Skeptics will want third-party evals and real deployment cost figures before declaring Inkling competitive. What to watch: benchmark results, license terms, and whether Thinking Machines can convert an impressive first model into a sustainable business.