Feature · Open models

'31' drew 4,900 likes: what this week's three open-model stories split over

"Scoring 31 on the AA Intelligence Index" drew about 4,900 likes. The same week Alibaba's Qwen got 13,700 with a three-word teaser, and Nous Research retrained a Chinese model on 'Western values.' We sorted the week's open-model news into three questions: score, weights and values.

1. Inkling 2 Small: 17B active out of 594B

   According to the Thinking Machines blog, Inkling 2 Small is a mixture-of-experts model with 594 billion total and 17 billion active parameters, accepting image input. Coding scores are 72.1% on DeepSWE v1.1, 86.3% on SWE-Bench Pro v2 (Hard) and 81.6% on Vibe Code Bench. The official post passed about 4,900 likes and 1.3 million views.

2. Reading the table: many models score higher on the index

   In the same blog's AA index v5.0 table, Claude Haiku 5.5 scores 36.5, DeepSeek-V4.1-Flash 32.6, GLM 5.3 31.7, Gemini 3.8 Flash 30.8, Inkling 2 Small 30.6, GLM-5.3-Flash 27.8 and Muse Spark Open 24.9. One X user called it the weakest major open release of the quarter and said to look at the price.

3. The point is the weights: Apache 2.0 next week

   For now it runs on Thinking Machines' Tinker without downloadable weights; the company says full weights will be released under Apache 2.0 next week after external safety testing, with other platforms in the coming weeks. It also listed limits: chat is less polished and it can pick poor tools in codebases with many skills.

4. Arete-27B: changing a Chinese model's values with 53 million tokens

   Nous Research retrained Qwen3.8-27B and says the 'propaganda rate' on its 803-prompt eval fell from 22% to 1%, with no significant loss on Hermes Bench. It released the eval, weights and training data. The figures come from an eval Nous designed itself and have not been independently verified.

5. MixedDavid Y. Tan 외 · X (Qwen) ↗

   Qwen 4: one teaser line and preview reports

   Alibaba's official post 'The becoming begins.' drew about 13,700 likes. One user who tried a preview said it looks on par with Opus 5.5 (about 550 likes); another replied that Opus 5.5 is not beatable anytime soon. Alibaba has not announced weights or official benchmarks for the preview.

6. What is left to confirm

   Thinking Machines' weights are a promise for next week, Qwen 4 has no price or official figures, and Arete's eval is self-designed. This week's reaction suggests the open-model contest is shifting from rank to whether weights are truly released and what values they carry.