Inkling 2 Mini released: League 60.1 and coding #4 — on a maker-reported coding score.
Thinking Machines Lab released Inkling 2 Mini on November 20, 2026 — a 26.4B-total, 3.4B-active looped MoE with open weights under Apache 2.0, in a single ~20GB checkpoint for local hardware. It enters our database with two tracked components — a leaderboard-sourced GPQA Diamond 86.6 and a maker-reported SWE-bench Verified 79.2% — which normalise to a League Score of 60.1 at 35% coverage: overall #36, but coding #4. This post works both normalisations in full and flags, explicitly, which of the two scores the maker reported itself.
By AI Pro League · Published

Inkling 2 Mini is the smallest model Thinking Machines Lab has released. Our November 21, 2026 changelog records it from the lab’s official Hugging Face model card (read today): released November 20, 2026 with open weights under Apache 2.0, a 26.4B-total / 3.4B-active multimodal looped mixture-of-experts — 24 physical layers, the middle 12 run three times for 48 layer passes, 5.4B effective active parameters per token — in a single ~20GB NVFP4/MXFP8 checkpoint built to run on local hardware, with a context window of up to 1M tokens and text and image inputs. It joins a database of 68 models, 56 of them scored, in the AI model rankings. What follows is computed from the database at build time.
The Inkling 2 Mini record, worked in full
Mini is the first Inkling to enter the database with two tracked components, and the two carry different kinds of evidence. GPQA Diamond carries 15% of the League Score: its raw 86.6 comes from the BenchLM GPQA leaderboard (measured Nov 21, 2026), an independent leaderboard source, and it normalises against the 28-model tracked field, which runs from 73.0 to 97.0 — worked step by step: (86.6 − 73.0) ÷ 24.0 × 100 = 56.7. SWE-bench Verified carries 20% of the League Score: its raw 79.2% (bash only) is maker-reported — it comes from the lab’s own model card (measured Nov 20, 2026), not from a leaderboard this site tracks — and it normalises against the 14-model tracked field, which runs from 51.0 to 96.0: (79.2 − 51.0) ÷ 45.0 × 100 = 62.7. How each board works is set out in our GPQA Diamond explainer and our SWE-bench Verified explainer; the standings themselves live on the GPQA Diamond leaderboard and the SWE-bench Verified leaderboard.
| Component | Raw | Field | Normalised | League weight | Source | Measured |
|---|---|---|---|---|---|---|
| SWE-bench Verified (bash only) | 79.2% | 51.0–96.0 (14 models) | 62.7 | 20% | Thinking Machines Lab model card — maker-reported | Nov 20, 2026 |
| GPQA Diamond | 86.6 | 73.0–97.0 (28 models) | 56.7 | 15% | BenchLM GPQA leaderboard | Nov 21, 2026 |
The League Score is the weighted sum over the coverage the model actually has: 62.7 × 20% + 56.7 × 15% = 21.05, divided by the 35% coverage gives 60.1. That places Mini at #36 overall, #24 on reasoning and #10 among open-weight models — and, because SWE-bench Verified is one of its two components, #4 on coding, a rank none of its higher-scoring siblings holds. Our score coverage explainer sets out why we print the coverage beside every score on this site.
Coding #4 — and what that rank rests on
| Coding rank | Model | SWE-bench raw | Normalised | League Score | Overall rank | Coverage |
|---|---|---|---|---|---|---|
| 1 | Claude Opus 5 | 96.0 | 100.0 | 93.1 | #5 | 90% |
| 2 | Claude Sonnet 5 | 85.2 | 76.0 | 77.0 | #18 | 75% |
| 3 | Kimi K2.6 | 80.2 | 64.9 | 68.3 | #28 | 35% |
| 4 | Inkling 2 Mini (new) | 79.2 | 62.7 | 60.1 | #36 | 35% |
| 5 | Mistral Medium 3.5 | 77.6 | 59.1 | 59.1 | #39 | 20% |
| 6 | Qwen3.5-397B-A17B | 76.2 | 56.0 | 59.5 | #38 | 35% |
Read the rank with its record attached. Mini’s #4 rests on a single coding component at 35% total coverage — and that component is the maker-reported one. The models around it on the coding table carry leaderboard-sourced SWE-bench records; Mini’s 79.2% has not yet been re-measured by a source independent of its maker. The rank is real arithmetic on the record we have; it is not yet a finding of the same grade as its neighbours’. The full coding table lives among our best AI models for coding.
The family ladder: the lowest GPQA, the second component
| Model | Released | GPQA raw | Normalised | League Score | Overall rank | Components | Coverage |
|---|---|---|---|---|---|---|---|
| Inkling | Jul 15, 2026 | 87.9 | 62.1 | 34.9 | #48 | 3 | 70% |
| Inkling Small | Jul 31, 2026 | 89.5 | 68.8 | 68.8 | #26 | 1 | 15% |
| Inkling 2 Small | Nov 5, 2026 | 93.0 | 83.3 | 83.3 | #13 | 1 | 15% |
| Inkling 2 Mini (new) | Nov 20, 2026 | 86.6 | 56.7 | 60.1 | #36 | 2 | 35% |
The ladder runs the wrong way for the newest model: Inkling at 87.9, Inkling Small at 89.5, Inkling 2 Small at 93.0 — and Mini, the family’s smallest and newest, at 86.6, the lowest GPQA Diamond in the family. Looping adds computation, not stored knowledge, and the GPQA record shows it. What Mini gains instead is breadth: it is the first family member whose record is not GPQA alone — Inkling 2 Small still leads the family, and the open table, on a single component at 15% coverage, while Mini’s second component takes its coverage to 35% and gives it the coding rank the rest of the family lacks. The family’s open-weight context is in our best open-weight models analysis and on the open-source models page.
Why there is no price in this post
Every priced model on this site gets a standard workload cost — 10M input + 2M output tokens at its recorded rates. Mini gets none, deliberately: its pricing record is null. It is an open-weights model — the cost of running it is the cost of the local hardware its ~20GB checkpoint fits on, not a per-token rate — and the Tinker rates for hosted access are platform rates, not model pricing, the same treatment the rest of the Inkling family receives in our database. A workload figure computed from Tinker’s list would price the platform, not the model, so this post does not print one. The rate cards we do record sit in our AI model pricing index.
What the League Score does not yet prove
- One of its two components is maker-reported. The SWE-bench Verified 79.2% (bash only) comes from Thinking Machines Lab’s own model card; unlike Mini’s GPQA component — and unlike the leaderboard-sourced components behind most scores on this site — it has not been published by an independent leaderboard. If a tracked leaderboard re-measures it lower, both the League Score and the coding rank move.
- It is 35% coverage. Three of the five components the League scores — LiveBench, MMLU-Pro and LMArena — have no Mini result. The 60.1 is a promising rumour with two data points, not a finding on a full record.
- Normalisation moves with the field. Both normalised values are computed against today’s tracked fields; a new maximum on either board re-scales Mini’s score without Mini being re-measured at all.
- The family comparison cuts both ways. Mini’s GPQA is the family’s lowest even as its record is the family’s second-deepest — breadth and peak are different claims, and this post makes only the breadth one.
The score to watch next is a leaderboard-sourced SWE-bench Verified: the first independent re-measurement of the 79.2% decides whether coding #4 is a finding or a launch-day figure. The model’s full record lives on its Inkling 2 Mini page, and its predecessor’s on the Inkling 2 Small page.