Inkling 2 Small released: a 594B MoE that enters the League unranked — for now.
Inkling 2 Small, Thinking Machines Lab’s second-generation model, was released on November 5, 2026 and enters our database today: a 594B-total, 17B-active mixture-of-experts model with image inputs, available only through the lab’s Tinker platform, with its weights still closed and no League Score yet. This post sets out what actually shipped, why the model is unranked, how it compares with the two earlier Inklings in our own records, and the dated weights promise to watch.
By AI Pro League · Published
Inkling 2 Small is the newest entry in our model database — the 60th model we track — added today from Thinking Machines Lab’s official announcement, recorded in our AI lab release tracker. The model page is Inkling 2 Small; the lab’s own account is the official Inkling 2 Small announcement, and everything below is either from that announcement, as recorded in our changelog, or computed from our own model records at build time.
What Inkling 2 Small is: the specs on record
Per the announcement, as recorded in our changelog: Inkling 2 Small is a mixture-of-experts model with 594 billion total parameters, of which 17 billion are active on a given token — an active share of just 2.9%, the MoE bargain in one number: frontier-scale capacity, small-model serving cost per token. It takes text and image inputs and returns text, and it is served through Tinker — the lab’s fine-tuning, sampling and Playground platform — at a 256K-token served context; our record carries the exact window as 262,144 tokens (262.1K). There is no separate public API listing in our database: the record’s only provider is Tinker.
Why Inkling 2 Small has no League Score yet
The launch tables report no result on any of the five benchmarks the League scores — no LiveBench, SWE-bench Verified, GPQA Diamond, MMLU-Pro or LMArena figure from a tracked leaderboard source. The headline number in the announcement, 31, is the lab’s Artificial Analysis Intelligence Index result, which the League does not score, so it cannot be converted into a League Score without inventing a mapping — and we do not invent mappings. The model therefore enters unranked at 0% coverage, exactly as our score coverage explainer describes for launch-day models. It is not alone: 15 of the 60 models we track currently carry no League Score, against 45 scored models in the AI model rankings. Inkling Small and Gemini 4 Carbon entered the same way, and both are still waiting for a tracked leaderboard to publish.
The Inkling family, side by side — our own data cut
Thinking Machines now fields three models in our database, and the family table is the honest comparison: the first-generation Inkling (released Jul 15, 2026) is the only scored Inkling, at 24.6 on 55% coverage — and it is also the open one, its weights public under Apache 2.0. Both Small models are unranked, and all three share the same 262.1K-token window. The generation change is in the weights column, not the window:
| Model | Released | Context | Inputs | Open weights | Licence | League Score |
|---|---|---|---|---|---|---|
| Inkling | Jul 15, 2026 | 262.1K | text, image, audio | Yes | Apache 2.0 | 24.6 | Unranked | Unranked |
Closed at launch, Apache 2.0 promised: the weights timeline
The sharpest break with the first generation is the launch posture: Inkling and Inkling Small shipped as open weights under Apache 2.0, while Inkling 2 Small launches closed. The closure is stated to be temporary — per the announcement recorded in our changelog, the lab says it intends to release the full weights under Apache 2.0 the following week, after external safety testing and red-teaming completes. Our Upcoming tracker carries the promise as a dated item — “Inkling 2 Small open weights”, earliest November 9 and latest November 15, 2026, conditional on that testing — in the Upcoming in AI tracker. It is the same API-first, weights-on-a-date pattern we tracked for the Kimi K3.1 Pro open weights release and for Mistral Large 4 — a pattern worth watching precisely because the last two labs kept the date.
What it costs — and why we record no price
Our record carries no pricing for Inkling 2 Small, and that is deliberate. The official Tinker docs list serverless-inference rates of $0.30 input / $1.00 output per million tokens (cached input $0.03), but our changelog entry records the decision not to file those as model pricing: serverless inference there is beta/waitlist, and neither earlier Inkling carries a price in our database either — the family’s recorded convention is unchanged. The full field, with the prices we do stand behind, is on our AI model pricing index. When a stable, public rate card exists, the record will gain one the way every other priced model did: from the primary source, on the record, with the check date printed.
What to watch next
- The weights window, Nov 9–15. Whether the Apache 2.0 release lands inside the promised week — and whether fine-tuned checkpoints become downloadable with it, as the tracker entry notes.
- The first tracked benchmark. A LiveBench or LMArena publication is what moves this model from unranked to ranked; our head-to-head pages already frame the wait — see Command A+ vs Inkling 2 Small compared.
- A stable price. A public, non-waitlist Tinker rate card would give the family its first recorded price in our database and put the model into the value rankings it currently cannot enter.
Until one of those lands, the accurate one-line summary is the unglamorous one: a very large, very sparsely-active model you can fine-tune today, cannot download yet, and cannot rank honestly until a tracked leaderboard measures it. The Thinking Machines Lab profile keeps the lab’s full release history in one place.