Inkling 2 Small open weights are out — and the first tracked score makes it the #1 open model.
The Inkling 2 Small open weights release landed on November 12, 2026 — seven days after the model launched on Tinker with its weights closed — on Hugging Face under Apache 2.0, inside the lab’s promised November 9–15 window. Its first tracked score is already on record: a GPQA Diamond 93.0 that normalises to a League Score of 87.0, overall #10 and #1 among open-weight models — on a single component, at 15% coverage. This post works the normalisation maths in full and states what that ranking does and does not prove.
By AI Pro League · Published

Inkling 2 Small open weights are the follow-through on the launch promise. Our November 13, 2026 changelog records the release from Thinking Machines Lab’s official Hugging Face model card (thinkingmachines/Inkling-2-Small, read November 13, 2026), and our Upcoming tracker item graduated the same day: shipped November 12, 2026. The model — Inkling 2 Small — entered the database on November 6, unranked and closed; our Inkling 2 Small release analysis set out that record, and this post is its sequel. Everything below is computed from the database at build time.
Inkling 2 Small open weights: what shipped on November 12
| Measure | Inkling 2 Small |
|---|---|
| Model release (Tinker, weights closed) | Nov 5, 2026 |
| Open weights released | November 12, 2026 — Hugging Face, Apache 2.0 |
| Checkpoints | NVFP4 main (25 safetensors files) + MXFP8 alternative |
| Architecture (launch record) | 594B total / 17B active MoE (2.9% active) |
| Context window on record / inputs | 262,144 tokens / Text, Image |
| GPQA Diamond (BenchLM, 2026-11-10) | 93.0 (#9 of 25 tracked) |
| League Score / ranks | 87.0 — #10 overall, #1 open, #9 reasoning |
| Score coverage | 15% — one component (GPQA Diamond) |
| Pricing on record | None (see below) |
The timeline is the first finding. At launch the lab said it intended to release the full weights under Apache 2.0 the following week, after external safety testing and red-teaming; our Upcoming in AI tracker carried the promise as a dated item, earliest November 9 and latest November 15, 2026. The weights landed on November 12 — inside the window, three days before its end. The record behind the promise is the same pattern we tracked for the Kimi K3.1 Pro open weights release: a dated weights commitment, kept on schedule.
The first tracked score: GPQA Diamond 93.0
The score predates the weights by two days. Our November 10, 2026 changelog records a GPQA Diamond 93.0 from the BenchLM GPQA leaderboard (measured November 10, 2026) — the model’s first result on any of the five leaderboards the League scores. On that board it sits #9 of the 25 models we track with a GPQA record, in a field running from 73.0 to 96.0 with a median of 90.9. Two distinctions matter. The lab’s own model card does not report a GPQA figure at all, and its headline index — 31 in the launch post, 30.6 on the card — is the Artificial Analysis Intelligence Index, which the League does not score. The 93.0 on this page is a tracked-leaderboard measurement, not a launch-table number. How the benchmark works is set out in our GPQA Diamond explainer, and the full field sits on the GPQA Diamond leaderboard.
How 93.0 becomes a League Score of 87.0
With one recorded component, the maths is checkable by hand. Normalisation stretches the recorded GPQA field — 25 models, raw scores from 73.0 to 96.0 — across a 0–100 scale: (93.0 − 73.0) ÷ 23.0 × 100 = 87.0. GPQA Diamond carries 15% of the League Score, and with no other component recorded the League Score is that normalised value: 87.0. That places the model at #10 of 51 scored models (64 tracked), 13.0 points behind the leader, Claude Fable 5.2 — and it makes the coverage column, not the score column, the honest read of the table below:
| Rank | Model | League Score | Coverage | Maker |
|---|---|---|---|---|
| 8 | GPT-5.6 Sol | 88.8 | 70% | OpenAI |
| 9 | GPT-6.1 Terra | 88.4 | 40% | OpenAI |
| 10 | Inkling 2 Small (new at #1 open) | 87.0 | 15% | Thinking Machines Lab |
| 11 | Gemini 4 Argon | 86.3 | 40% | Google DeepMind |
| 12 | Claude Sonnet 5.5 | 85.0 | 40% | Anthropic |
Every neighbour in that band carries at least 40% coverage. Inkling 2 Small’s 15% is the thinnest record in the top 12 — a point the standings themselves, in the AI model rankings, print beside the score.
#1 among open-weight models
The open title is the headline rank: #1 of the 23 scored models among 31 open-weight models we track, ahead of DeepSeek V4.1 Flash at 84.4 and Kimi K3.1 Pro at 75.4. Read it with the coverage column attached: the runner-up’s 84.4 rests on 55% coverage — more than three times the record behind the 87.0 it trails. The weights release makes the title more than a label: since November 12 the #1 open model in our table is one whose checkpoints anyone can download, inspect and fine-tune under Apache 2.0. The field as a whole is ranked in our best open-weight models analysis and on the open-weight models leaderboard; the head-to-head frame is DeepSeek V4.1 Flash vs Inkling 2 Small compared.
The Inkling family on one benchmark
All three Inklings now carry a GPQA record from the same leaderboard, which makes the family table a clean generational read — and a coverage lesson in the same rows:
| Model | Released | GPQA raw | Normalised | League Score (rank) | Coverage |
|---|---|---|---|---|---|
| Inkling | Jul 15, 2026 | 87.9 | 64.8 | 36.1 (#43) | 70% |
| Inkling Small | Jul 31, 2026 | 89.5 | 71.7 | 71.7 (#22) | 15% |
| Inkling 2 Small | Nov 5, 2026 | 93.0 | 87.0 | 87.0 (#10) | 15% |
The raw line climbs steadily — 87.9 to 89.5 to 93.0 — yet the first-generation Inkling scores only 36.1 overall, because its deeper record (70% coverage across three components) averages the strong GPQA component against weaker LiveBench and LMArena components. Inkling Small mirrors the new model: the same 15% single-component shape, one generation earlier.
Why we still record no price
Our record carries no pricing for Inkling 2 Small, unchanged from launch. Tinker lists serverless rates for the model, but our changelog records the standing decision not to file them as model pricing: they are one platform’s serverless-inference rates, and no Inkling carries a recorded price in our database. The weights release does not change that: a downloadable checkpoint has no per-token price, only the hosting bill of whoever serves it. The prices we do stand behind are on our AI model pricing index; when a stable public rate card exists, the record will gain one.
What the #1 open ranking does not yet prove
- It is one component. Coverage is 15% — GPQA Diamond only, the joint-thinnest class in the League (6 of 51 scored models sit at 15%; the field median is 40%, and 26 of 51 scores rest on a single component). No LiveBench, SWE-bench Verified, MMLU-Pro or LMArena result is recorded yet, so 87.0 is a finding about graduate-level science reasoning — it is also #9 on reasoning — not about coding, agentic work or breadth. A promising rumour vs a finding: treat the rank as the rumour until a second component lands, the standing caveat in our score coverage explainer.
- A second component can pull it down. The family table above is the precedent: the first Inkling’s deeper record averages its GPQA component down to 36.1. If Inkling 2 Small’s next tracked component normalises below 87.0, the composite falls — and the open title, held by 2.6 points over a rival with 55% coverage, is the first thing at risk.
- Open weights is not small weights. The downloadable checkpoint is the full 594B-total model in NVFP4 or MXFP8 form, per the model card — “Small” names the 17B active footprint per token, not the download. The licence is genuinely permissive (Apache 2.0, as recorded); the hardware question is separate, and it is a cluster question, not a laptop one.
None of that dims the week itself: a dated weights promise kept inside its window, and a first tracked score that puts a downloadable model at the top of the open table. The score to watch is the second one — whichever tracked leaderboard reports on Inkling 2 Small next will test whether 87.0 is a ceiling or a starting point. The model’s full record lives on its Inkling 2 Small page, the lab’s history on the Thinking Machines Lab profile, and the standings themselves in the AI model rankings.