Qwen3.8 27B
on 2× AMD RX 7600 XT 16GB · llama.cpp · 35,000 ctx
Summary
User reports Qwen3.8 27B at 550 t/s prompt processing and 30 t/s generation on an RX 7600 XT 16GB plus RX 9060 XT 16GB, in a moderate context of 35k tokens.
Setup is llama.cpp (latest git) with UD-Q4_K_XL GGUF weights, f16 KV cache, tensor split 60/40, flash attention on, and draft-MTP plus ngram-mod speculative decoding. ROCm was considerably faster than Vulkan.
Adaptive KV streaming reached over 800 t/s prefill but limited the user to lower quantizations and left the second card unused. The user asks whether Strata or another fork would do better on this 32GB RAM system.
Compare this run
Medians of plain runs: one device, one request, no speculative decoding, the whole model in the device's memory. The report above is one setup, so read it beside them and not as one of them.
Qwen3.8 27B at 4-bit by hardware
- NVIDIA RTX 509032 GB80.0 t/s · 7 runs
- 67.0 t/s · one run
- NVIDIA RTX 5070 Ti16 GB54.7 t/s · one run
- 50.0 t/s · one run
- NVIDIA RTX 409024 GB40.0 t/s · 3 runs
- AMD RX 7900 XTX24 GB40.0 t/s · 7 runs
Bars are medians on one scale, 0 to 94.0 tokens/s. Each dot is one run and opens its report. Press ▶ to watch text arrive at a row's speed.
Models most run on the AMD RX 7600 XT 16GB
- Qwen3.8 27B3-bit18.0 t/s · one run
Bars are medians on one scale, 0 to 18.0 tokens/s. Press ▶ to watch text arrive at a row's speed.
Everything reported on the AMD RX 7600 XT 16GB →Every Qwen3.8 report →Estimate Qwen3.8 on this card →