Qwen falls apart at tool-calling while gemma4 shines in a DIY agentic loop
spammmmmmmmy · reddit · 2026-09-23
A Reddit user shares hands-on experience building a simple DIY agentic loop: the model is told its available tools, proposes commands with a RUN: prefix, and the user approves or edits them each turn.
Key findings:
- Go-to models Qwen3.6 35B and Qwen3.8 27B could not follow instructions consistently, spiraling into insane feedback loops once their context clogged — catting whole files instead of using ast-grep, and leaking raw thinking tokens and template delimiters into output.
- As a last resort the user tried gemma4:12b and found it surprisingly obedient and accurate; gemma4:26b was even better at planning work and proposing next steps.
- The author also backtested Qwen2.5:3b-instruct, reportedly trained specifically for tool-calling.
They conclude there may be a trade-off slider between "creative" and "obedient" behavior, and open a discussion on which models handle tool-calling best.
More from coding & agent
- MatBrain splits reasoning from tool use: two models screen 30,000 crystal candidates in 48 hours — bravo_abad · 2026-09-23
- Firecrawl Raises $75M Series B, Launches Alexandria Knowledge Library for AI Agents — omarsar0 · 2026-09-23
- Five AI agents bypassed a permissions broker in ten minutes using 'start' instead of 'stop' — TrifleHopeful5418 · 2026-09-23
- Shopify CEO who pushed staff to use AI now 'horrified' — the 'Slop Grenades' story — srchvrs · 2026-09-23
- Lovable adds Claude Opus 5.5 and GPT-6 Sol, auto-routing between frontier models — AlexandrePesant · 2026-09-23
- Dev laments agents built around KV caches, wants inference-first chips — dbreunig · 2026-09-23