Thinking Machines ships Inkling 2 Mini: 26B looped transformer, top open model under 100B
ChowdhuryNeil · x · 2026-11-21
Thinking Machines (Mira Murati's lab) released Inkling 2 Mini, an open-weights model answering complaints that Inkling 2 Small (594B) was too big.
- Architecture: looped transformer — middle layers run three times per token, enabling more internal reasoning with fewer output tokens
- Specs: 26.4B total params (3.4B active), 20GB memory, runs on a laptop
- Benchmarks: 48.2% on DeepSWE v1.1 and 46.5% on Vals Index v2.2, claimed best among open-weights models under 100B; AA Index of 20.1, suited for agentic coding, faster than similarly performing models
Boosters call it an impressive small open-weights release; charts show it beating Qwen3.8-27B and Gemma 4 in its class.
More from coding & agent
- Route agents by task type, not fixed rules: not every request needs a frontier LLM call — rohanpaul_ai · 2026-11-21
- OpenClaw's Skill Workshop 2.0 lets your agent review and improve its own skills — heyneighbor · 2026-11-21
- Sentry CEO: coding agents nearly ate 1TB of disk in a week, he's now more against handoff — zeeg · 2026-11-21
- Perplexity Computer comes to Chromium browsers via extension, working across tabs and tools — perplexity_ai · 2026-11-21
- VibeArmory 0.3.0 adds on-device ML to monitor Mac AI agent activity locally — agf2007 · 2026-11-21
- AI-powered research pipeline cuts customer feedback cycle from 10 days to under 2 — pavandavuluri · 2026-11-21