Thinking Machines open-sources variable-loop research model: loops chosen at inference, up to 79.6% on SWEBench Verified
danielhanchen · x · 2026-11-21
Thinking Machines Lab released TML-d24-varloop-research (Apache 2.0), a variable-looped counterpart to Inkling 2 Mini: the number of passes through the middle 12 looped layers can be chosen at inference time on a single set of weights, enabling cleaner studies of how extra looped compute affects capability.
Architecture:
- 24-layer decoder-only transformer with sparse MoE (8 of 128 experts per token), hybrid local/global attention
- Natively multimodal: images encoded via a ViT and processed jointly with text
- 26.4B total params, 3.4B active at 1 pass; each extra pass adds 1.0B effective active params (5.4B at 3 passes, 7.4B at 5)
- Supports NVFP4/MXFP8 numerics
Benchmarks (mini-swe-agent v2.4.6): more loops clearly help agentic coding — SWEBench Verified climbs from 56.3% at 1 pass to 79.6% at 5, SWEBench Pro v2 (Hard) hits 78.9%, both beating fixed-3-pass Inkling 2 Mini (79.2% / 77.9%); DeepSWE v1.1 reaches 48.2%. Gains saturate after 3 passes.
More from Models
- Gemini 4 Neon fixes complex codebase issues nearly on par with Opus 5.5 — op7418 · 2026-11-21
- Google AI Pro users hit 0% weekly limits in Antigravity, refresh in 4 days — boyus · 2026-11-21
- Dev swaps GLM 5.3 Flash for Thinking Machines' Inkling 2 Small in Windows observability app — michellechen · 2026-11-21
- Building camera-motion measurement to diagnose open world model failures: measure before training — PrateekVJoshi · 2026-11-21
- User says ChatGPT's ultrafast mode is impossible to go back from — TheMoonMidas · 2026-11-21
- House report on PRC models urges national US open-weight AI initiative — evijit · 2026-11-21