15 Qwen3.8 27B variants tested on dual 5070 Ti: Tess-4.5 cuts coding time 40%
gamedevsam · reddit · 2026-11-08
The author benchmarked 15 community fine-tunes and quants of Qwen3.8 27B against the base UD-Q6K on 2x RTX 5070 Ti (32GB), with 100-question screening rounds and 1,300-question full runs (AA-LCR, MRCR, AA-Omniscience, BFCL v4 subset, SciCode via NeMo-Skills).
Results
- Overall: base Q6 63.3 > Tess-4.5 61.4 > 4-bit 60.8 > ThinkingCap 57.5 — no fine-tune beat base overall
- On a homemade test of 25 real bugs from the author's projects, Tess-4.5 matched base on fixes (18.3 vs 18.0) while using 40% fewer tokens (14K vs 23K per bug) — faster because it thinks less, not because it decodes faster (both 36 t/s)
- ThinkingCap thinks half as long but its extra correct fact answers came mostly from guessing (25 right out of 146 attempts on questions base declined)
Verdict: keep two models
- Base for chat and interactive work: won 42 of 200 BFCL multi-turn conversations Tess lost (Tess skips steps), declines/hedges on 69% of facts it gets wrong vs 56%
- Tess-4.5 for autonomous coding agents checked by tests — and surprisingly it was the more cautious one on the 100 software questions (42 declines vs 20)
Full settings (q40 KV cache, 262K window, 790 t/s prompt reading) and open-source harnesses are linked in the original write-up.
More from coding & agent
- Agent swarms show endogenous division of scientific labor across 1,800 sessions — _AustinCalvert_ · 2026-11-08
- Doodle Smash: hand-drawn doodles become 3D fighting characters, built in two days — chongdashu · 2026-11-08
- Lessons from building an MCP server whose tools cost real money — eminbudak · 2026-11-08
- Viggle Remake MCP adds a ComfyUI worker: int8 quantization runs on 32GB VRAM — init-5 · 2026-11-08
- Nous Research's Hermes agent can now phone you on WhatsApp when a scheduled job needs attention — Teknium · 2026-11-08
- Agent-Native Vork Engine Renders 24x Faster With a 2,500-Token API — LinusEkenstam · 2026-11-08