Inkling 2 Mini runs locally in 16GB RAM: 5.4B-active MoE tops its size class
simonguozirui · x · 2026-11-21
Unsloth shipped GGUF quantizations and a guide for Inkling 2 Mini, Thinking Machines' Apache 2.0 looped-transformer MoE (26.4B total / 5.4B active, 1M context) for coding, agentic and vision work. It scores 48.2% on DeepSWE v1.1 and 77.9% on SWE-bench Pro v2 (Hard) — best for its size, beating Qwen3.8-27B and even 501B Beam 1.0 on several coding boards. Hardware needs: 9GB RAM at 1-bit, 16GB-class at 2-3 bit, 24GB for lossless 8-bit; runs via Unsloth and llama.cpp.
More from coding & agent
- Inkling 2 Mini runs fully offline on a MacBook, builds an airplane simulator in minutes — LiTianleli · 2026-11-21
- Thinking Machines' Inkling 2 Mini: 26B looped transformer tops open-model coding boards — victormustar · 2026-11-21
- Replit engineer: 'I no longer need anything locally for my job' — amasad · 2026-11-21
- Replit Agent lands in Microsoft Teams to turn chats into action — amasad · 2026-11-21
- Dev shares a simple local Claude Code combo: use another frontier model as a second pair of eyes — DavidWells · 2026-11-21
- LangChain roadmap: every major player converging on persistent agents — hwchase17 · 2026-11-21