Qwen3.8-27B vs Qwen3.6-27B (vLLM, fp8, 256k ctx, bf16 kvcache) on a private benchmark

Agentic benchmark based on University of Pisa low-level operating systems exams: Italian instructions, C++/x86_64, custom kernel code, compilation + actually booting/testing in QEMU. Exact pass/fail, no partial credit. Same clean pi agent setup: Qwen3.8-27B: 6/10 — 60% Qwen3.6-27B: 3/10 — 30%