Qwen3.8 125B (6B active) Flash-Next
on 2× AMD RX 7900 XTX · Strata · 112 ctx
Use cases
codingagentic
Summary
User benchmarks Qwen3.8-Flash-Next at IQ3_XXS on two RX 7900 XTX cards, comparing engines with MTP drafting on and prompt cache off. The fastest row is the Strata-rdna3 fork with tensor parallel at 92 t/s decode on a 112-token prompt, 104 t/s at 3.5k and 99 t/s at 32k, with prompt reads of 2,619 and 3,030 t/s.
Setup is the Strata-rdna3 fork (fa5c621) on ROCm 7.2.4 with a Ryzen 7 5800X and 64 GB DDR4; the default config pins about 39 GiB of experts in RAM, and the user ran with --resident-budget-gib 24 --vram-reserve-mib 1536.
Avalon on one card reached 89/98/88 t/s decode but read prompts at about half the fork's speed, and its current main crashed on the 32k prompt. Upstream Strata 0.2.11 and 0.2.13.1 decoded at 76 and 75 t/s, llama.cpp b11815 at 33 t/s. In real agent runs the fork decoded at about 120 t/s, roughly 16% less wall time per run. A side note gives Qwen3.8 27B on llama.cpp at 49-56 t/s on one XTX and 68-73 t/s with experimental tensor split on two cards. About 74% of agent wall time is decode and 16% prompt reading.
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 125B · 6B active at 4-bit by hardware
- 129.2 t/s · one run
- 105.5 t/s · 6 runs
- 57.2 t/s · 2 runs
- 55.0 t/s · one run
- 36.7 t/s · one run
- NVIDIA DGX Spark128 GB29.8 t/s · 2 runs
Bars are medians on one scale, 0 to 232.0 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 7900 XTX
- Qwen3.8 27B4-bit40.0 t/s · 7 runs
- 142.8 t/s · one run
- Gemma 4 31B4-bit30.1 t/s · one run
- Gemma 4 8B16-bit58.0 t/s · one run
- 125.5 t/s · one run
Bars are medians on one scale, 0 to 142.8 tokens/s. Each dot is one run and opens its report. Press ▶ to watch text arrive at a row's speed.
Everything reported on the AMD RX 7900 XTX →Every Qwen3.8 report →Estimate Qwen3.8 on this card →