AI BriefingMoE rule fragility
Key point
In a 4x RTX 3090 test, MoE Qwen models violated strict bash rules more often than a dense 27B.
Details
Running Qwen3.5-27B dense, Qwen3.5-122B-A10B MoE, and Qwen3.6-35B-A3B MoE for 20+ real-world agent sessions each on a 4x RTX 3090 setup, the MoE models consistently showed weaker adherence to a strict bash allow-list.
The experiment was run on a multi-agent orchestrator assuming 1 to 6 concurrent OpenCode sessions, 30-60k token prompts, and 262k context. The allow rules were very strict, permitting only exact patterns like uv run scripts/<name>.py while banning shell decorators such as | head, | tail, timeout, 2>&1, absolute-path Read, and cd && ... chains.
- Qwen3.5-27B: dense, INT8 (AWQ-BF16-INT8), FP8 KV, MTP speculative decoding
- Qwen3.5-122B-A10B: MoE, AWQ-INT4, FP8 KV
- Qwen3.6-35B-A3B: MoE, FP8 weights, FP16 KV
The author aggregated Avg prompt throughput, Avg generation throughput, and Running: N reqs from vLLM v0.19.0 logs at 10-second intervals, with each cell computed as the average over the corresponding concurrency range. The table was measured at 250W, and the conclusion is that regardless of model size, active parameter count, or fine-tuning target, MoE models are worse at maintaining global rules than the dense 27B.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.