Inkling
Thinking Machines Lab’s first broadly released model, out 15 July 2026 as open weights: a 975B-parameter mixture-of-experts (41B active) reasoning natively over text, images, and audio, with a “thinking effort” dial exposed to the user. The lab — founded by former OpenAI CTO Mira Murati — had shipped only the Tinker fine-tuning API (Oct 2025) before this.
This page was created 19 Jul 2026, four days into the model’s public life — visibly incomplete by design. The local corpus ends before its release, so the tweet layer awaits a fresh pull; reception below is day-of press only. A dossier pass is pending.
Sources
Official
- 2026-07-15 Inkling: Our open-weights model — the release announcement.
- 2026-07-15 Inkling model card — sizes, training mix, the effort dial, and the lab’s own capability positioning.
Writing & commentary
- 2026-07-15 TechCrunch — frames the release as the lab’s “bet against one-size-fits-all AI.”
- 2026-07-15 VentureBeat — claims the release is “focused on low cost and ‘resistance to censorship.’”
- 2026-07-15 Bloomberg — “first AI model for broad use” framing; the lab-history angle.
- 2026-07-16 Simon Willison — day-after technical read.
Tweets
tk — corpus pass pending; the local dbs end before this model existed. Records will populate when tweets are cited.
Official record
- Released 15 July 2026 as open weights (downloadable via Hugging Face). The lab’s first released model; its first product was the Tinker fine-tuning API (Oct 2025).
- Architecture as published: 975B-parameter mixture-of-experts, ~41B active per token; 1M-token context window; trained on ~45T tokens spanning text, image, audio, and video; reasons natively over text, images, and audio.
- The effort dial: a user-set “thinking effort” parameter (0.2–0.99) trading answer quality against token cost — exposed as a first-class control rather than a hidden router.
- Inkling-Small previewed alongside: 276B parameters, ~12B active, similar training recipe; launch materials state it matches or beats the flagship on several benchmarks at far lower cost.
- The lab’s own positioning: launch materials are candid that Inkling is “not the strongest overall model available today, open or closed” — the pitch is token-efficiency and the best open-weights showing from a US lab, per day-of coverage. REPORTED
- tk — license terms, exact benchmark table, checkpoint naming, API/serving story.
History
- 2026-07-15 World at release: lands six days after OpenAI’s GPT-5.6 generation launch and into an open-weights field led from China (Kimi K2 lineage, DeepSeek, Qwen) — day-of coverage read it against exactly that field. The founding team’s OpenAI provenance (Murati, ex-CTO) carried much of the framing.
- tk — day-of reception beyond press; first independent evals; whether “resistance to censorship” (VentureBeat’s phrase) is the lab’s language or the reporter’s; naturalist reads as they emerge.
Impressions
tk — too fresh; no attributed character reads yet. The corpus pass and first-weeks discourse will fill this. Nothing is asserted in the meantime.