Qwen3.8 Flash Next IQ3_S hits 90 tk/s locally, low-poly 3D generation beats expectations
migsperez · reddit · 2026-10-13
Continuing his benchmark series on Opencode, the author ran Qwen3.8-Flash-Next IQ3S via Strata on an AMD R9700 32GB with 64GB RAM, achieving 90 tk/s decode — roughly 3x faster than the 27B model — despite expecting terrible results from a Q3 quant.
Key finding: standard prompts gave mediocre output, but asking for a plan first, then building it, and specifying "low poly style but with many polygons" produced a major leap — iterating this way usually makes other models worse. Example comparisons: IQ3S vs Qwen 3.8 Max. He also notes many mid-tier models performed surprisingly terribly on Opencode.
More from coding & agent
- AI Engineer New York opens with a print daily: 300K-dev telemetry says org, not model — jonmarkgo · 2026-10-13
- One-line decorator turns any combinatorial solver into a differentiable NN layer — PMinervini · 2026-10-13
- Google's enterprise Gemini agent: one memory across Workspace, M365 and Slack, routes work to Claude — BenBajarin · 2026-10-13
- Reverse-Engineering Elgato 4K Capture Card With T3 Code, Latency Down to ~24ms — talkaboutdesign · 2026-10-13
- Air Street Capital's Nathan Benaich: Codex predicted prompts are very good — nathanbenaich · 2026-10-13
- AI bot posting on Reddit for its user sparks debate on where agent approval lines should be — ChatGPTworkBot · 2026-10-13