Builds / RTX 5090

RTX 5090 (NixOS, Ollama 0.32.13): Qwen3.8 27B q4 at 124-149 tok/s, about 1.7-2x Qwen3.6 27B

Last updated: 4 Nov 2026

Source: dev.to ↗📏 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: not stated
Motherboard: not stated
RAM: not stated
PSU: not stated
Other: NixOS; Ollama 0.32.13 pinned via an overlay (bundled llama.cpp b10380); OLLAMA_NUM_PARALLEL=1
Reported total cost: not stated
Power: not stated

Measured speed (generation tokens/sec)

Model	Quant	Engine	tok/s	Prompt processing (tok/s)	Details
Qwen3.8 27B	Ollama qwen3.8:27b q4 (17 GB on disk)	Ollama 0.32.13	124	not stated	context: four agent-style prompts📏 Generation speed over the same four prompts on one box: 124-149 tok/s (RigBase lists the low end); 143 tok/s in the author's Ollama-vs-vLLM table. The post does not say whether Ollama used the model's built-in MTP layer; by RigBase's arithmetic, speeds above ~105 tok/s exceed one full read of a 17 GB file per token at the 5090's 1,792 GB/s spec bandwidth, which suggests drafting was active. The unofficial Unsloth GGUF on Ollama 0.32.9 gave 72 tok/s (28-token generation, coarse)
Qwen3.6 27B legacy	Ollama qwen3.6:27b (the post does not state the quant; Ollama's registry tag is Q4_K_M per localaimaster.com/benchmarks)	Ollama 0.32.13	72	not stated	context: four agent-style prompts📏 Same box, runtime and four prompts: 72-73 tok/s (low end listed)

Posted 3 Nov 2026 on DEV Community (originally on wada-dev.com). VRAM for Qwen3.8 27B q4: 19.7 GB at 32K context, 23.6 GB at 128K, 29.2 GB at 256K. vLLM with the NVFP4 weights only started in eager mode (163,840 context) and generated 27.8 tok/s with concurrency 1.01x, so it is not listed as a result. Muse Glimmer 30B ran at 75 tok/s on the same setup (quant not stated, not listed). On a JSON-only prompt Qwen3.8 finished thinking in 2.7 s vs 29.3 s for Qwen3.6. CPU, RAM and quant details beyond 'q4' are not given in the post.

Current GPU street price (estimate, not the builder's cost)

1× RTX 5090: ≈€7,140 each (≈$8,003), new, approx. price as of 2026-10-24, lowest listing on billiger.de (MSI GeForce RTX 5090 (lowest variant))

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://dev.to/kohei_wada/qwen38-27b-measured-on-an-rtx-5090-and-why-vllm-didnt-work-2m14

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-nixos-ollama-qwen38-27b-kohei-wada.svg)](https://getrigbase.com/builds/rtx-5090-nixos-ollama-qwen38-27b-kohei-wada/)

HTML

<a href="https://getrigbase.com/builds/rtx-5090-nixos-ollama-qwen38-27b-kohei-wada/" rel="noopener"><img src="https://getrigbase.com/badge/rtx-5090-nixos-ollama-qwen38-27b-kohei-wada.svg" alt="Measured on RigBase" height="20"/></a>

Preview: · 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 Qwen3.6 27B? · Can RTX 5090 run Qwen3.8 27B?