Lobotomized weights in practice: Q8 Qwen3.6 beat hyped QFN iq3_s as a coding agent
Ok-Doughnut-3022 · reddit · 2026-11-03
A Reddit user compares Strata-hosted QFN3.8 iq3s against the smaller Qwen3.6 35B Q8 and finds the more hyped QFN less usable as a coding agent.
- Even with recommended temp/top-k settings to curb overthinking and doomloops, QFN iq3 remained unreliable — "dunning-kruger with illusions of grandeur," while Qwen3.6, though overly literal and needing back-and-forth clarification, reliably executed tasks like "a talented but unmotivated programmer."
- The gap comes from precision: QFN iq3 is heavily quantized while Qwen3.6 runs at Q8. The author questions whether mid-size MoEs stay meaningful once they require lobotomization to load, and wonders if Qwen 4 will even matter locally.
The post pushes back on the community's focus on PP/decode speeds over real-world usability of aggressive quantizations.
More from coding & agent
- OpenAI's new Git backend Pando cuts push latency ~80%, serves ~1PB of Git data daily — OpenAIDevs · 2026-11-03
- OpenAI engineer details Pando, the source-control rewrite built for the agent era — OpenAIDevs · 2026-11-03
- Trajectory launches post-training platform with GS1 and Continual Graders research previews — simonguozirui · 2026-11-03
- SMBs are rebuilding their entire GTM stack with agent-native tools, says new thread — garrytan · 2026-11-03
- Compacting Claude Code at 400K instead of 967K cut this dev's cost per request by 43% — Awkward-Let-4628 · 2026-11-03
- Capy 0.7.0 ships Captains orchestration GA; YC's Garry Tan reports 5X issue-fixing throughput — garrytan · 2026-11-03