Qwen 3.8 27B fails complex coding task, lags behind DeepSeek and GLM
myreala · reddit · 2026-08-19
A user reported a disappointing experience with Qwen 3.8 27B on a complex C kernel coding task. The objective was to adjust thread count limits in a TTS execution kernel originally written by DeepSeek Pro. Despite running on a single 3090 with 150k context and fp8 kv cache for about six hours across multiple sessions, Qwen entered a loop, produced unusable code, and suffered from tool call failures. In contrast, GLM 5.3 completed the entire task in about twenty minutes. The user expressed reluctance to invest in higher quantization or hardware given the poor results.
More from Models
- Grok Bot Ran a 12-Hour Data Analysis Nonstop, Handling API Rate Limits Itself — Daniel_Farinax · 2026-10-16
- One year of Recursive Language Models: GPT-5-mini beats GPT-5 by 110% on 132k-token tasks, cheaper too — a1zhang · 2026-10-16
- Users Report Claude Models Suddenly More Compliant After Forced Compaction — RealSharpNinja · 2026-10-16
- Prediction: Opus 5.5-level models on laptops by 2027, backed by Qwen 27B data — cephaloform · 2026-10-16
- Users Say Krea 2 Output Degraded Sharply Since Krea 3 Beta Launched — szywis · 2026-10-16
- LLM-as-a-Verifier lifts Nemotron 3.5 Super by 9-10 points on agentic benchmarks — simonguozirui · 2026-10-16