Inkling: Our open-weights model (Willison)
Simon Willison‘s 2026-07-16 note on Thinking Machines Lab‘s Inkling release — the source that routed here.
What it says
- The release facts: 975B total / 41B active MoE, 45T tokens of text/image/audio/video, Apache-2.0, with Inkling-Small (276B/12B) still in testing.
- Quotes the lab’s own framing — “a good open-weights base for customization: multimodal capabilities, efficient thinking, and availability on Tinker for fine-tuning” — and its concession that Inkling is “not the strongest overall model available today, open or closed.”
- Flags the thin training-data documentation: public internet content plus “publicly accessible data repositories” and third-party sets, no breakdown.
- The ecosystem line: “It’s good to see the US open weights ecosystem gain a new viable contender to join NVIDIA Nemotron and Gemma 4.”
- Access is through the Tinker platform’s API — OpenAI-compatible chat-completions, bearer-token auth, demonstrated with curl for both text and image input. Another data point for the compatibility-as-strategy thread (synthesis): a brand-new lab still clones OpenAI‘s API shape.
The pelican
Willison ran his usual SVG pelican on a bicycle probe, then fed the drawing back to the model to describe. It called its own bird a “stork or seagull” — a “cheerful, flat-vector cartoon illustration featuring a white bird riding a bicycle.” Funny, and mildly informative: the vision path reads the image fine, it just doesn’t recognize what the generation path was aiming at.
Tier & provenance
T2 — a methodology-transparent independent practitioner relaying first-party facts, not first-party
documentation. The harder numbers on inkling (architecture, benchmark table, VRAM) come from the
lab’s own announcement and model card (T1), which this post links. freshness: volatile — a release-day snapshot.
Related
inkling · thinking-machines-lab · simon-willison · inkling-announcement · open-weight-models · tinker