Edition of Saturday 21 November 2026

Inkling 2 Mini

Comments on “Inkling 2 Mini”

Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked

Thinking Machines Lab released Inkling 2 Mini, an open-weights looped transformer: a 26.4B-parameter mixture-of-experts model that runs its middle layers three times per token. The ~20GB footprint fits a laptop, and vLLM supports it from day one. The company reports the top open-weights scores under 100B parameters on DeepSWE v1.1 (48.2%) and Vals Index v2.2 (46.5%); these figures are self-reported, and setups vary by model. It trails on multi-turn chat and constraint-following. A variable-loop checkpoint supports research on looping and chain-of-thought monitorability.

Read the original at thinkingmachines.ai. The judges saw only its text, with the site, the author and link domains removed; they are shown again after ranking.

Why it ranked

No. 18 of 118 with a score of 0.69. The panel rated it highest on new information (0.93) and technical depth (0.92), lowest on audience breadth (0.66). Placed No. 18, where the score alone would also have put it: the fifth AI & machine learning story of the day (−0.040). The placement terms move stories; they are not part of the score.

AI judgeJudge 2claude-haiku-5-5
Open-weights model release with detailed architecture, benchmarks and looping/monitorability research; substantive technical detail, though vendor-authored with self-reported benchmarks and a narrow-to-moderate audience.
AI judgeJudge 3gpt-5.6-luna
A technically detailed open-weight model release introduces looped-transformer architecture, low-bit local inference, extensive benchmark results, training details, and monitorability experiments. Its claims are substantially evidenced, though launch framing and benchmark comparisons warrant modest overclaim uncertainty.

Each judge wrote its line after answering the rubric, with the site and the author still hidden. Every answer behind this is in the table under Score.

Score

0.69total under rubric v2.3

Judges’ mean answer per question, on a scale from 0 to 1. Questions marked minus count against the score.
  • Significance0.69
  • Reader interest0.80
  • Audience breadth0.66
  • Practical relevance0.76
  • Substantiated0.86
  • Technical depth0.92
  • New information0.93
  • Counts against: Promotional0.20
  • Counts against: Overclaim0.35

Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).

QuestionJudge 1Judge 2Judge 3MeanSpreadWeightContribution (weight times mean)
SignificanceHow much does this change what people in tech know or can do?0.710.550.820.690.27+0.24+0.167
Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on?0.700.800.900.800.20+0.26+0.208
Audience breadthHow many people who build technology does it concern?0.670.600.720.660.12+0.13+0.086
Practical relevanceHow useful is it to someone who builds software or tech products?0.770.700.800.760.10+0.22+0.167
SubstantiatedHow well is it backed? Consider whether it is the primary source, and whether its main claims come with data, code, a demo, a reproducible method or citations.0.960.750.880.860.21+0.05+0.043
Technical depthHow technically substantive is it?0.950.850.960.920.11+0.05+0.046
New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)?0.910.900.990.930.09+0.05+0.047
PromotionalIs it mainly marketing with little substance (a press release, ad or sponsored post with no real information)? A launch post that carries real news is not promotional.0.110.300.180.200.19−0.19−0.037
OverclaimDo the headline or conclusions claim more than the body supports?0.540.300.220.350.32−0.11−0.039
Sum of contributions+0.687
Claim-check penalty0.000
Total, as published0.687

Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.

The badge for this ranking, for the author's page or anyone's: one line of HTML, and it links back here.

Ranked #18 on Rubric, 21 Nov 2026
<a href="https://rubric.hrih.si/editions/2026-11-21/18/"><img src="https://rubric.hrih.si/badge/2026-11-21/18.svg" alt="Ranked #18 on Rubric, 21 Nov 2026" width="245" height="28"></a>

Claim check

Claim check: 8 no source given. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.

  1. No source givenBenchmark result

    Inkling 2 Mini leads every open-weights model with less than 100B total parameters on benchmarks such as DeepSWE v1.1 (48.2%) and Vals Index v2.2 (46.5%).

    The article gives no link or citation for it.

    The article says

    “Inkling 2 Mini leads every open-weights model with less than 100B total parameters on benchmarks such as DeepSWE v1.1 (48.2%) and Vals Index v2.2 (46.5%”

    The source says

    No source to compare with.

  2. No source givenNumber

    Inkling 2 Mini is a Mixture-of-Experts model with 26.4B total parameters, 3.4B active, and 5.4B weight activations per token to account for looping.

    The article gives no link or citation for it.

    The article says

    “Inkling 2 Mini is a Mixture-of-Experts model with 26.4B total parameters, 3.4B active, and 5.4B weight activations per token to account for looping”

    The source says

    No source to compare with.

  3. No source givenNumber

    The variable-loop model's no-CoT time horizon improved from 16 seconds at one pass to 37 seconds at three passes.

    The article gives no link or citation for it.

    The article says

    “The no-CoT time horizon of TML-d24-varloop-research rises from 16 seconds at one pass to 37 seconds at three”

    The source says

    No source to compare with.

  4. No source givenBenchmark result

    Inkling 2 Mini scored 0.703 on intelligence per watt on NVIDIA DGX Spark, versus 0.430 for Qwen3.6-35B-A3B, 0.159 for Gemma-4-26B-A4B and 0.036 for Qwen3.5-27B.

    The article gives no link or citation for it.

    The article says

    “Inkling 2 Mini scored 0.703 running on NVIDIA DGX Spark, significantly higher than other models: 0.430 for Qwen3.6-35B-A3B, 0.159 for Gemma-4-26B-A4B and 0.036 for Qwen3.5-27B”

    The source says

    No source to compare with.

  5. No source givenNumber

    The total memory footprint of Inkling 2 Mini is about 20GB, allowing it to run on consumer laptops and desktop AI computers.

    The article gives no link or citation for it.

    The article says

    “The total memory footprint of Inkling 2 Mini is ~20GB, allowing it to run on local devices such as consumer laptops and desktop AI computers”

    The source says

    No source to compare with.

  6. No source givenNumber

    The non-looping model cost 7.5% less compute to train but required 140% more inference compute to match the looped version's loss on 512k-token agentic traces.

    The article gives no link or citation for it.

    The article says

    “The non-looping model cost 7.5% less compute to train, but required 140% more inference compute to match the looped version’s loss on 512k-token agentic traces”

    The source says

    No source to compare with.

  7. No source givenNumber

    Reasoning at 5 passes was less than half as long as at 1 pass (1,661 vs. 3,703 tokens on average) on GPQA Diamond questions answered correctly at all pass counts, with no loss of legibility.

    The article gives no link or citation for it.

    The article says

    “Reasoning at 5 passes was less than half as long as at 1 pass (1,661 vs. 3,703 tokens on average), but no less legible”

    The source says

    No source to compare with.

  8. No source givenBenchmark result

    Among CoTs of 750–1,250 tokens, Inkling 2 Mini and the variable-loop model follow CoT-control instructions at most 0.1% of the time, versus 0.3% for Inkling 2 Small, 0.4% for Inkling-Small and 1.3% for Qwen3.8-27B.

    The article gives no link or citation for it.

    The article says

    “follow the instruction at most 0.1% of the time, compared with 0.3% for Inkling 2 Small, 0.4% for Inkling-Small, and 1.3% for Qwen3.8-27B”

    The source says

    No source to compare with.

Also covered by

Elsewhere

Elsewhere: HN 850 pts, front page

Discussion on Hacker News

Recorded for context only, checked 21 Nov 2026, 05:47 UTC. It played no part in the ranking.

Panel for this edition

Four AI agents ranked this edition: three judges scored every story blind to its site and author, and a fourth compared the top stories’ claims with the sources they cite. Summaries are written by AI.

Model versions are pinned for every edition. How ranking works describes rubric v2.3.

Comments

Readers’ comments on this story have a page of their own, ranked by a public rubric rather than by votes.

Comments on “Inkling 2 Mini”

More on Rubric

From the last 7 editions, chosen by shared topic, score and date.