Thinking Machines Lab (Inkling)
Thinking Machines Lab (Inkling)
One-line summary: mira-murati's US lab; its first model Inkling (975B-parameter open-weight MoE) is pitched as a Western alternative to Chinese open weights, betting on customization / fine-tuning over leaderboard dominance.
What it is
Founded by mira-murati (former OpenAI CTO) after one of the largest seed rounds ever. Its first release, Inkling, is an open-weight foundation model announced ~2026-07-16: a mixture-of-experts model with 975B total parameters, ~41B active, trained on 45T tokens of text/image/audio/video and reasoning natively across all four modalities, ~1M-token context. Reuters framed it as "a Western alternative to the Chinese open weight models Deepseek and Quen." Per alexander-wissner-gross in 2026-07-17-podcast-moonshots-mira-murati-s-975b-open-model-ramin-hasani-on, the evals put it stronger than Nvidia's Nemotron but weaker than GLM 5.2 — a real Western frontier-open-weight entrant, not the strongest overall.
Why it matters to artificial-intelligence
Inkling is the flagship of the contrarian "customization over frontier" bet: Murati explicitly is not claiming the best model on earth; the business is fine-tuning-as-a-service so enterprises adapt an owned, cheaper open model on-prem. ramin-hasani: "their business is fine tuning... they're deliberately leaving some room for fine tuning" — potentially "one to two orders of magnitude more tokens" on the customization side. The tension (alexander-wissner-gross): OpenAI already tried reinforcement-fine-tuning-as-a-service and shut it down; Thinking Machines is betting RFT is a durable paradigm, which may or may not hold. It's a datapoint for open-vs-closed-source-model-economics and the sovereignty/on-prem thesis in ai-safety-as-regulatory-capture. Dave Blundin's read: that a top ex-OpenAI founder built around this thesis is itself a signal the open-weight-catch-up path is viewed as viable inside the best labs.
Key facts
- Inkling: 975B total / ~41B active (MoE), 45T tokens, natively multimodal, ~1M context, open-weight, on-prem fine-tunable.
- Eval position: > Nvidia Nemotron, < GLM 5.2 (per AWG).
- Founder: Mira Murati (ex-OpenAI CTO); very large seed round.
- Strategy: customization / fine-tuning-as-a-service, not leaderboard dominance.
Open questions
- Is the RFT/fine-tuning-as-a-service bet durable, or does a sufficiently generalist base model kill the need for fine-tuning? (AWG's doubt.)
- Does a US open-weight tier matter more if the US restricts Chinese open weights? (See ai-self-regulatory-body.)
Sources
- 2026-07-17-podcast-moonshots-mira-murati-s-975b-open-model-ramin-hasani-on (multi-context, vault/sources/)
- 2026-07-18-podcast-all-in-podcast-can-the-ai-industry-regulate-itself-stripe-wants (multi-context, vault/sources/) — Sacks on Inkling as a fine-tuning platform for cheaper open models.