VX27 · llm-lab

Benchmark reports

Local LLM head-to-head reports from a 2x RTX PRO 4000 Blackwell rig. Frozen eval sets, 3-run medians with warmup discard, quant-matched comparisons, every row names its winner.

• Qwen3.6-35B vs Qwen3.8-27B Exhaustive campaign, August 2026. Agentic suite, tool calling, quick-quality, speed, VRAM, power, KV-cache and quant arms, with an Opus 4.6 frontier reference. Verdict: keep Qwen3.6-35B.

← vx27.com