MiMo v2.6 Distill flops 0/10 as a coding agent on a 16GB Mac, losing to plain Qwen 9B
GapNew4766 · reddit · 2026-11-18
Testing on a 16GB M4 MacBook Air with llama.cpp, the author ran three small models as coding agents on the same task (CSV to HTML spending report with SVG chart), 15-minute limit, 10 automated tests. Qwen3.5-4B scored 7/10, Qwen3.5-9B 8/10 (10/10 at 32k ctx), while Xiaomi's MiMo v2.6 Distill — which Xiaomi claims lifts SWE-Bench Pro from 32 to 44.6 — went 0/10 three times, timing out after writing 278 lines with a syntax error it couldn't recover from, and crashing on TypeErrors even with thinking on. Full setup, prompt, CSV and checker are shared for reproduction; author discloses he works on atomic agent.
More from coding & agent
- dottxt tool-call index shows LFM3-1.7B broken tool calls jump from 7.19% to 15.96% across providers — remilouf · 2026-11-18
- Pocket Folio adds API key + CLI so AI agents can auto-fill your card binders — cnakazawa · 2026-11-18
- Merge Adds Triggers to Agent Handler: Agents Fire Automatically on Emails and PRs — shensi · 2026-11-18
- Frontier models race jets through the Top Gun canyon: Astra wins, Opus refuses to fly — abhi3188 · 2026-11-18
- Running a literature research project inside Grok bot, from question to grant aim — MaxUnfried · 2026-11-18
- Incredibuild spins up Islo, an enterprise 'software factory' for AI agents, with $15M from Insight Partners — Shruti_0810 · 2026-11-18