Qwen3.8 27B hallucinated a whole feature: 120k tokens in, it never read the plan
KingCpzombie · reddit · 2026-09-03
- A Redditor reports Qwen3.8 27B Q8 (Unsloth quant, llama.cpp) going badly off-script in a coding agent workflow: 120k tokens in, the model had never read the plan and implemented a totally different, never-mentioned feature.
- The author admits 3.8's reliability made him lazy — he reviewed the plan and final output but not the execution process; the recommended temperature of 1 and lack of process monitoring are the likely culprits.
- Full config included: llama-server, 200k context, dual Vulkan GPUs, temp 1 / top-p 0.95 / top-k 20, draft-mtp speculative decoding.
More from coding & agent
- Hamel Husain on decision models for evals: don't trust the 0.5 threshold default — HamelHusain · 2026-10-21
- HF Staff Open-Sources Self-Iterating 3D Asset Prompt With Vision Critique Loops for Small Quantized Models — mervenoyann · 2026-10-21
- Open models below 4-bit struggle at self-iterating 3D asset generation, HF dev finds — mervenoyann · 2026-10-21
- Developer Moves AI Agents to a Dedicated Linux Workstation Using Incus System Containers — kevinkern · 2026-10-21
- Motion Launches a General Video Agent That Edits Footage and Mimics Any Reference Style — mhdfaran · 2026-10-21
- MCP Platform NitroStack Wins Product Hunt Product of the Day, Focuses on Retention — EstablishmentFun4373 · 2026-10-21