Joakimpalm-Zen/granite-4.0-h-small-runner-report
Preview • Updated • 54
Third-party files run on Xyntetik Runner and compared with llama.cpp and the publisher's reference, each bound to the exact file by SHA-256.
Note IBM's Q8_0 (granitehybrid, Mamba-2 hybrid MoE): 3/5 prompts byte-identical to llama.cpp, envelope manifest included.
Note Community Q4_0 (nemotron_h_moe): 4/5 prompts byte-identical to llama.cpp, envelope manifest included.
Note 0.6B draft / 8B target acceptance across draft depths: 44.1% at k=2, and why tok/round must not be tuned on.
Note The nemotron_h Q8_0 Runner quantised itself, 5/6 greedy-identical vs llama.cpp.