Agent-tuned MiMo scores 0/10 in a real loop while base Qwen 3.5 gets 7-8/10
GapNew4766 · reddit · 2026-11-18
A hands-on benchmark in a real agent loop (write a CSV-to-HTML spending report script, 15-min limit, 16GB MacBook, atomic agent harness):
- Qwen 3.5 4B: 7/10 — opened its generated page, caught its own reversed top-5 table and fixed it
- Qwen 3.5 9B: 8/10 — one-shot draft, only missed SVG charts
- Xiaomi's agent-tuned MiMo v2.6 distill 9B: 0/10 timeout — broken first draft with a syntax error it couldn't fix from the error message, ending by writing another script-length reply; with thinking on, it crashed 6 times on the same category/amount mix-up, 5 of 11 edits changed nothing
Takeaway: in a real agent loop, reading an error and making the right fix matters far more than an "agent-tuned" label. Setup: llama.cpp, q4km, 64k ctx, temp 0.2; prompt, CSV and checker are public.
More from coding & agent
- Fable 5.2 turns a fake mobile game ad into a real playable game: 70 min, 2,982 lines, $29.52 — Dan_Jeffries1 · 2026-11-18
- Google-Backed Kaggle Agent Contest Offers $400K to Crack Engineering Benchmark PLE — amankhan · 2026-11-18
- Dev snark: before running 10 agents, ask if you could even build software with one terminal — KevinNaughtonJr · 2026-11-18
- Fan-built zero-socket MCP bridge gives Project Zomboid an AI radio Game Master — Triminum · 2026-11-18
- Suno confirms its MCP includes visuals generation — suno · 2026-11-18
- Oxlint adopts tsrs backend: type-aware linting up to 7x faster, 75% less memory — cnakazawa · 2026-11-18