Qwen3.6-35B on a 16GB Mac beats GPT-5-mini on honest tool-calling, at 20x the latency
Mission-Sir-1614 · reddit · 2026-10-22
A developer staged a tool-calling showdown between Qwen3.6-35B-A3B (streamed from SSD on a 16GB Mac mini M4 via Kraits, $0/run) and OpenAI's gpt-5-mini API, refereed by a neutral CrewAI agent with all traffic logged.
- Easy ticket: both went 3/3; gpt-5-mini took 13-18s per run vs 63-72s.
- Trap-laden ticket (duplicate customers, deprecated tools, currency traps, 503 retry, endless pending): Kraits finished 3/3 with zero traps hit; gpt-5-mini executed every action correctly and avoided all traps but promised in 2 of 3 runs to send a follow-up email despite having no email tool. Kraits plainly said it couldn't send email in all three runs.
- Thinking mode is what keeps it honest: with thinking off, Kraits dropped to 2/3, with one run making zero tool calls and hallucinating a full fake completion.
- The cost: 9.7-14.1 minutes per run vs 25-35 seconds — but free and fully local.
- MoE X-ray: expert activation logs show deep layers (32-39) diverge most during tool calls; layer 34 expert 108 fires on 73-78% of tool-call words.
More from coding & agent
- Couple builds 2.5D pixel MMO 'Lumoria' with Claude Opus — first boss-fight gameplay revealed — AkiraVero · 2026-10-22
- Proof of Agent network upgrade: hosted mining paused 24-72h, agent ops playbook shared — adamamcbride · 2026-10-22
- Agent ops tips: retire bots with clean handoff docs and cap retries at 1-3 loops — alexcovo_eth · 2026-10-22
- Supabase ships unified .NET SDK v8: all packages in one repo, one version number — dshukertjr · 2026-10-22
- MCP Incident Room: An Interactive TypeScript Prototype for Learning MCP Design — kilmex33 · 2026-10-22
- Perplexity brings Graphify to Computer: visualize any codebase as an interactive map, 5,000 free credits per user — AravSrinivas · 2026-10-22