3.6B TwIL-LM3-Pro runs locally on 4GB VRAM, claims 35% lead over VibeThinker-3B
glenbeer · x · 2026-10-04
webAI released TwIL-LM3-Pro, a 3.6B-parameter open-source model built for local inference: the Q4 build is just 2.09 GiB and runs on CPU or 4GB of VRAM, with no cloud or per-query API costs.
Vendor-reported benchmarks:
- Highest recorded headline score in formal logic among compared small models, beating VibeThinker-3B by 35%, Qwen3.5-4B by 24%, and Liquid AI's LFM2.5-8B-A1B by 47%
- 95% on SVAMP and 64.1% on MuSR, also claimed as best-in-class for small models
- The open-source family hit 500K downloads in its first month
The release echoes Karpathy's thesis that the next big step in AI comes from tiny models packing dense intelligence. Note all scores are self-reported evaluations.
More from Models
- OpenAI's DevDay dots agent called underwhelming in hands-on comparisons vs Grok Bot — alexisgallagher · 2026-10-05
- Which Qwen3.8-27B fine-tune is best? Comparing ThinkingCap, Swift 1.5 and QwenPi — Sam Witteveen · 2026-10-05
- Switching models mid-session is a broken experience, users complain — Gauri_the_great · 2026-10-05
- Reddit user suspects Sonnet 5.5 burns through usage limits suspiciously fast — paulofilip3 · 2026-10-04
- Claude storage map: new Pro/Max sessions go cloud-only starting Oct 6 — BenSimonDev · 2026-10-04
- Grok 4.7 tops Artificial Analysis' Cyber Index, beating Claude Opus 5.5 and ChatGPT 6 Astra — XFreeze · 2026-10-04