CALIBRATED 2026-07-27 · REC 003
Local AI Frontier

FRONTIER LAB

Lab notes & dead ends

The open notebook behind the published benchmarks: methodology experiments, failed approaches, and the things we tried that didn't work. Publishing dead ends is part of the contract.

Recent experiments

  • BM-0032026-07-27vulkan
    Qwen3-Coder-30B-A3B quantization sweep: Q4_K_M vs. Q6_K vs. Q8_0 on Arc B60 Pro

    On 24GB Arc, Q6_K is the sweet spot: quality recovers to within 2 points of Q8_0 on HumanEval while staying 38% faster and fitting comfortably. Q8_0 barely fits and leaves no room for context. Q4_K_M is the budget pick when VRAM is tight.

  • BM-0022026-07-26cuda
    Qwen3-Coder-30B-A3B: NVIDIA CUDA vs. Arc Vulkan at Q4_K_M

    CUDA on the NVIDIA reference produces 52.1 tok/s vs. 38.6 tok/s on Arc Vulkan — a 35% generation-speed lead at the same 24GB VRAM tier. If your workload is latency-sensitive and CUDA is an option, NVIDIA still wins. Arc remains the better $/tok for budget builds.

  • BM-0012026-07-25vulkan
    Qwen3-Coder-30B-A3B on Arc B60 Pro vs. dual RX 9700

    Vulkan on Intel Battlemage lands within 12% of dual-RX 9700 for code generation at Q4_K_M, at roughly half the system cost. Arc is the price-performance leader for sub-$2K builds; multi-GPU AMD wins raw throughput once you accept the complexity.

Documented failures