More
Best rig by budgetRent before you buySubmit a buildNewsletterFree datasetAdvertise
Submit build

Builds / RTX 5090

RTX 5090 + 128 GB DDR5 + Optane: Qwen3.8 Flash-Next Q4 to Q8 at 104.0 to 43.2 tok/s (Strata, ngram table on SSD)

Last updated:

Source: toolhalla.ai ↗📏 How measured: reported by the original source; RigBase re-checked the numbers at that source
GPU(s)
RTX 5090
Amazon · eBay (used) · Newegg
VRAM / unified memory (GB)
32
CPU
AMD Ryzen 9 9950X3D (16 cores / 32 threads)
Amazon · eBay (used) · Newegg
Motherboard
not stated
RAM
128 GB DDR5-5600 (4x32 GB)
Amazon · eBay (used) · Newegg
PSU
not stated
Other
Ubuntu 24.04.5, driver 595.99.02; Strata 0.1.48 (CUDA); 400 GB Intel Optane P5800X holds the ngram/PLE lookup table and the MTP helper; other shards on Optane (Q4, Q8) or Samsung 980 Pro NVMe (Q5, Q6); experts beyond VRAM in system RAM
Reported total cost
not stated
Power
not stated

Measured speed (generation tokens/sec)

ModelQuantEnginetok/sPrompt processing (tok/s)Details
Qwen3.8 Flash-NextUnsloth UD-Q4_K_XL (111.3 GB incl. 28.8 GB IQ4_NL ngram table)Strata 0.1.48 (CUDA)104.01,856 @ 2,048-token prompt
est. TTFT @8k ≈ 4.4 s (calculated from measured prefill)		⚡ spec-decode (MTP + lookup drafting (--spec 4))context: 2,048-token prompt, depth 0📏 Toolhalla lab notes, 27-28 Oct 2026: 2,048-token prompt after context seeding, 128 output tokens, one client, mean of three requests after warm-up, timed from a separate Mac over LAN (llama-benchy e9be344); Strata flags --spec 4 --spec-min-p 0.5 --suffix-draft 3 --lookup-chain 4, int8 KV, --resident-budget-gib 71, ngram/PLE table on Optane P5800X (--ple-io direct); 96.5 at 32K depth, 96.8 at 100K
Qwen3.8 Flash-NextUnsloth UD-Q5_K_XL (158.3 GB incl. 54.4 GB Q8_0 ngram table)Strata 0.1.48 (CUDA)81.81,336 @ 2,048-token prompt
est. TTFT @8k ≈ 6.1 s (calculated from measured prefill)		⚡ spec-decode (MTP + lookup drafting (--spec 4))context: 2,048-token prompt, depth 0📏 Toolhalla lab notes, 27-28 Oct 2026: 2,048-token prompt after context seeding, 128 output tokens, one client, mean of three requests after warm-up, timed from a separate Mac over LAN (llama-benchy e9be344); Strata flags --spec 4 --spec-min-p 0.5 --suffix-draft 3 --lookup-chain 4, int8 KV, --resident-budget-gib 71, ngram/PLE table on Optane P5800X (--ple-io direct); 76.6 at 32K depth, 70.0 at 100K
Qwen3.8 Flash-NextUnsloth UD-Q6_K_XL (169.2 GB incl. 54.4 GB Q8_0 ngram table)Strata 0.1.48 (CUDA)60.6869 @ 2,048-token prompt
est. TTFT @8k ≈ 9.4 s (calculated from measured prefill)		⚡ spec-decode (MTP + lookup drafting (--spec 4))context: 2,048-token prompt, depth 0📏 Toolhalla lab notes, 27-28 Oct 2026: 2,048-token prompt after context seeding, 128 output tokens, one client, mean of three requests after warm-up, timed from a separate Mac over LAN (llama-benchy e9be344); Strata flags --spec 4 --spec-min-p 0.5 --suffix-draft 3 --lookup-chain 4, int8 KV, --resident-budget-gib 71, ngram/PLE table on Optane P5800X (--ple-io direct); 62.7 at 32K depth, 56.7 at 100K
Qwen3.8 Flash-NextQ8_0 (188.2 GB incl. 54.4 GB Q8_0 ngram table)Strata 0.1.48 (CUDA)43.2517 @ 2,048-token prompt
est. TTFT @8k ≈ 15.9 s (calculated from measured prefill)		⚡ spec-decode (MTP + lookup drafting (--spec 4))context: 2,048-token prompt, depth 0📏 Toolhalla lab notes, 27-28 Oct 2026: 2,048-token prompt after context seeding, 128 output tokens, one client, mean of three requests after warm-up, timed from a separate Mac over LAN (llama-benchy e9be344); Strata flags --spec 4 --spec-min-p 0.5 --suffix-draft 3 --lookup-chain 4, int8 KV, --resident-budget-gib 71, ngram/PLE table on Optane P5800X (--ple-io direct); 42.9 at 32K depth, 43.0 at 100K

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.

Toolhalla lab notes, published 28 Oct 2026 (measured 27-28 Oct). One workstation, five quants run one after another on the same 28-cell grid; speculative drafting was on for every run, and outputs were capped at 128 tokens, so these are short-reply speeds. The UD-Q3_K_XL run (158.6 tok/s) is below Q4 and not listed. Throughput only: answer quality was not measured.

Current GPU street price (estimate, not the builder's cost)Estimated GPU cost for this build today: ≈€7,140
New retail prices in Germany incl. VAT; used cards are often much cheaper. Prices change daily.
Source: https://toolhalla.ai/blog/qwen3-8-q3-q4-q5-q6-q8-ngram-optane-rtx-5090

Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.

Embed this result

Copy into a README or blog. Links back to this build. See badge docs (CC BY 4.0 attribution). The SVG does not invent speeds — numbers stay on the build page.

Markdown

[![Measured on RigBase](https://getrigbase.com/badge/rtx-5090-ryzen-9950x3d-128gb-optane-qwen38-flash-next-toolhalla.svg)](https://getrigbase.com/builds/rtx-5090-ryzen-9950x3d-128gb-optane-qwen38-flash-next-toolhalla/)

HTML

<a href="https://getrigbase.com/builds/rtx-5090-ryzen-9950x3d-128gb-optane-qwen38-flash-next-toolhalla/" rel="noopener"><img src="https://getrigbase.com/badge/rtx-5090-ryzen-9950x3d-128gb-optane-qwen38-flash-next-toolhalla.svg" alt="Measured on RigBase" height="20"/></a>

Preview: Measured on RigBase · How to use

Rent before you buy: try this model on a cloud GPU for a few dollars first: Runpod (referral link; sign up with Google to get the new-user bonus)

Quick answers: Can RTX 5090 run Qwen3.8 Flash-Next?