Use cases
agenticcodinglong-context
Summary
User benchmarks Qwen3.8 Flash-Next (MoE, ~78-82 GiB of weights) on an RTX 5070 Ti 16 GB with 64 GB DDR5, reporting 89.2 t/s decode on the final tuned Strata 0.1.59 stack.
Setup is Strata with an IQ3_S pack, experts streamed from pinned system RAM, 65536-token KV-resident window; the same run measured 109.2 t/s at 8k context, 37.3k t/s hot prefill and 57.7 ms TTFT.
Basalt 0.3.3 reached 101.0 t/s decode and 54.5k t/s prefill but returned empty answers at the RAM edge, so production stayed on Strata; Unsloth managed 27.3 t/s out of the box and 47.9 t/s after tuning on the same GGUF. Decode falls to 86-89 t/s at 180-200k context, and a warm DFlash2 probe hit 111 t/s.
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 NVIDIA RTX 5070 Ti
- 137.7 t/s · one run
- Qwen3.8 27B4-bit54.7 t/s · one run
- Qwen3.8 27B3-bit53.0 t/s · one run
Bars are medians on one scale, 0 to 137.7 tokens/s. Press ▶ to watch text arrive at a row's speed.
Everything reported on the NVIDIA RTX 5070 Ti →Every Qwen3.8 report →Estimate Qwen3.8 on this card →