Qwopus3.8 27B Coder-Flash
2× AMD Radeon AI PRO R9700 32GB · llama.cpp · 262,144 ctx
- generation:
- 26.5-68.4 tokens/s
- prompt processing (prefill):
- 269-1009 tokens/s
- quant:
- Q5_K_M (GGUF)
- kv:
- q8_0
User reports a 10.4-hour unattended agentic-coding run with Qwopus3.8-27B-Coder-Flash on a Radeon AI PRO R9700 32GB and Tiel-Coder-35B-A3B-MTP on an RX 9070 XT 16GB, generating 549K tokens from 46.8M prompt tokens. Setup is llama.cpp (Vulkan for the 27B at Q5_K_M with 262,144 context, ROCm 10.2 for the MoE at UD-IQ3_XXS with 196,608 context), q8_0 KV cache, MTP speculative decoding with up to 3 draft tokens, all layers on GPU. Overall generation was 41.1 t/s for the 27B and 43.1 t/s for the MoE; median per-request generation was 44.6 t/s (26.5 to 68.4) and 58.4 t/s (21.5 to 102.1). Overall prompt processing was 656 t/s and 616 t/s, with medians of 660 t/s (269 to 1,009) and 711 t/s (374 to 868). MTP draft acceptance averaged 0.689 and 0.831. The 27B used about 28.5 GB of the R9700's 32 GB at 262K context and stayed usable at about 30 t/s at 190K context. In a separate short-context benchmark the 27B ran about 69 t/s on Vulkan versus about 61 t/s on ROCm, and the MoE about 135 t/s on ROCm versus about 63 t/s on Vulkan.