DeepSeek V4.1 Flash on llama.cpp: long-context slowdown diagnosed and mitigated

Running DeepSeek V4.1 Flash locally via a llama.cpp PR started at 27 tok/s and dropped to 11 tok/s at 50k tokens; after a fix it only falls to 24 tok/s, though agent use still fails, so the author will try Unsloth next.

きしだൠ(K1S)

@kisArchived 10 days agoOriginally posted 2026.10.21