Running Qwen3-Flash-Next locally for coding: 90GB quantized weights, prefill pain at 7-8M tokens
intermundia · reddit · 2026-10-19
A developer shares a local long-coding-session setup: serving a quantized Qwen 3 Flash Next (90GB weights) through LM Studio, with homemade tweaks to manage memory and context, on a customized deepseek harness 0.2.1-alpha.1.
His takeaways: the model codes well but overthinks relentlessly ("but wait.. but what if.."), and thinking appears to be strictly on/off with no dial; prefill also becomes a serious pain once the session ingests around 7-8 million tokens. He asks if anyone has found a way to tone down the model's thinking.
More from coding & agent
- Open-source LichtFeld Studio unifies 3DGS training, editing and MCP automation — janusch_patas · 2026-10-19
- CLI-Anything hits 54.9k GitHub stars by auto-generating CLIs to make all software agent-native — bibryam · 2026-10-19
- Dev raves about vibe coding with Opus 5.5: 'like a high-IQ friend who gets everything right' — marclou · 2026-10-19
- Jev founder's harness engineering PDF claims 200x faster, 400x cheaper coding agents — blaizedsouza · 2026-10-19
- A literature-review workflow with Claude, Althea MCP, Lacuna and Overleaf — hugo_larochelle · 2026-10-19
- Headroom: open-source context compression cuts coding agent tokens by 20% — headroomlabs-ai · 2026-10-19