Builds / RTX 5090 (400 W) + RTX 5060 Ti 16GB (180 W, peer expert tier
RTX 5090 + RTX 5060 Ti, 128 GB RAM: Qwen3.8 Flash-Next on Basalt 0.2.0 at 307 tok/s prose (NVFP4), 342 at UD-IQ3_XXS
Last updated:
- GPU(s)
- RTX 5090 (400 W) + RTX 5060 Ti 16GB (180 W, peer expert tier)
Amazon · eBay (used) · Newegg - VRAM / unified memory (GB)
- 48
- CPU
- not stated
- Motherboard
- not stated
- RAM
- 128 GB
Amazon · eBay (used) · Newegg - PSU
- not stated
- Other
- Linux, Basalt 0.2.0 (MIT, a fork of Strata), int8 KV, experts that do not fit in VRAM in system RAM; MTP drafting on by default (output identical to plain decoding)
- Reported total cost
- not stated
- Power
- RTX 5090 capped at 400 W, RTX 5060 Ti at 180 W
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| Qwen3.8 Flash-Next | NVFP4-DQ6_K Basalt pack (NVIDIA NVFP4 experts, Q6_K dense, 122.0 GiB) | Basalt 0.2.0 | 307 | 7,807 @ 16K-token prompt est. TTFT @8k ≈ 1.0 s (calculated from measured prefill) | ⚡ spec-decode (MTP)context: single stream, prose📏 Engine author's best measured single-stream speed: prose 307, structured output 601, chat 335 tok/s; numbers move with every release |
| Qwen3.8 Flash-Next | Unsloth UD-IQ4_XS Basalt pack (90.9 GiB) | Basalt 0.2.0 | 300 | not stated | ⚡ spec-decode (MTP)context: single stream, prose📏 Best measured: prose 300, structured 608, chat 363 tok/s |
| Qwen3.8 Flash-Next | Unsloth UD-IQ3_XXS Basalt pack (81.6 GiB) | Basalt 0.2.0 | 342 | not stated | ⚡ spec-decode (MTP)context: single stream, prose📏 Best measured: prose 342, structured 658, chat 399 tok/s |
| Qwen3.8 Flash-Next | Q8_0 Basalt pack (Unsloth, 178.3 GiB) | Basalt 0.2.0 | 112 | not stated | ⚡ spec-decode (MTP)context: single stream, prose📏 Best measured: prose 112, structured 166, chat 101 tok/s (most experts in RAM) |
Prompt processing = prefill speed (how fast the prompt is read) as reported by the source, with the prompt length it was measured at; blank = the source does not report it. est. TTFT @8k is calculated from the measured prefill speed (8,192 ÷ prefill tok/s), shown only when prefill was measured on a prompt of 512+ tokens. It is not a measurement and ignores load time and caching. measured TTFT is the source's own time-to-first-token. Why prefill matters for coding assistants.
Engine author's own numbers (jesdga95, Hugging Face pack card updated 8 Nov 2026; precision docs on GitHub). These are best measured speeds, not averages. Quality from the same card (100 wikitext-2 chunks vs a llama.cpp BF16 reference): NVFP4-DQ6_K KLD 0.0624 / 91.65% same top-1, UD-IQ4_XS 0.0818 / 90.52%, UD-IQ3_XXS 0.1188 / 88.67%, Q8_0 0.0200 / 95.34% (llama.cpp's Q8_0 scores 0.0325 / 93.98%).
https://github.com/jesdga95/basalt/blob/main/docs/precision/qwen38-flash-next.md
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.
Speculative-decoding rows on this page are not like-for-like with plain decode. See how much MTP and DFlash really speed things up (measured pairs).