Qwen3.8 35B (3B active)
on P104-100 · llama.cpp
Use cases
coding
Summary
User reports Qwen3.8-35B-A3B at about 14.5 t/s generation on a P104-100 8GB with 24GB DDR3 system RAM.
Setup is llama.cpp turboquant with IQ2_M GGUF; prompt processing ran about 75-81 t/s.
User also tried Qwen3.8 27B at Q1_M and Q3_XSS (slow), Qwythos 9B Q4 (overthought), Qwen3.8 9B Q8 (fast but poor quality) and LFM2.5-8B-A1B Q4_K_M (decent), and liked the Qwen3.8 35B-A3B most. They want 20-25 t/s at 256k context for Java/C++/Rust coding.