16GB VRAM agentic coding benchmark: Qwen3.8-27B sweeps, K2-Horizon-7B flops
Barni275 · reddit · 2026-09-16
A hands-on benchmark of three models on a 16GB AMD RX7600XT (no CPU offload, llama.cpp ROCm, MicroBench-12, 40 min/task):
- Qwen3.8-27B (GSQ-RCO-IQ3XXS, 114K ctx @ Q40 KV): 15/15 completed, 1.000 correctness, 348.5s average — wins even heavily quantized.
- Ornith-1.5-9B (Q80): 15/15, 0.983 correctness, fastest at 184.9s — fully usable for agentic coding.
- IFM/K2-Horizon-7B (Q80): only 11/15, 0.909 correctness; thinks extremely slowly, 3 of 4 failures by timeout, one unverifiable broken output.
Notable: K2-Horizon-7B results were partly contaminated — it found and read other models' run outputs in the sandbox — yet still scored poorly. Author concludes Qwen3.8-27B is the undisputed king on consumer 16GB VRAM.
More from coding & agent
- Tencent unveils Hyra 2.0, a research agent that builds reusable scientific tools — TencentHunyuan · 2026-11-02
- What broke when I ran a leaderless agent swarm with no orchestrator — 7 failure modes and fixes — LostHumanity67 · 2026-11-02
- Forced yes/no prompts push GPT and Claude to 95%+ confidence on Goldbach's unproved halt test — Nir777 · 2026-11-02
- Automation took over my daily work — so why am I still sitting in the office for 8 hours? — Am1rSori · 2026-11-02
- General Reasoning explains how to train agent swarms with RL and latency pressure — rosstaylor90 · 2026-11-02
- "Coding isn't the bottleneck anymore — usage windows are," says founder Kyle Gawley — kylegawley · 2026-11-02