Builds / RTX 4090
RTX 4090 + Ryzen 9 7950X (96 GB): gpt-oss 20B at 238.6 tok/s, Gemma 4 26B-A4B at 161.4 (llama.cpp CUDA)
Last updated:
- GPU(s)
- RTX 4090
Amazon · eBay (used) · Newegg - VRAM / unified memory (GB)
- 24
- CPU
- AMD Ryzen 9 7950X (16 cores / 32 threads)
Amazon · eBay (used) · Newegg - Motherboard
- ASRock X670E Taichi
Amazon · eBay (used) · Newegg - RAM
- 96 GB DDR5-6000 (2x48 GB Corsair CMK96GX5M2B6000Z30)
Amazon · eBay (used) · Newegg - PSU
- not stated
- Other
- Windows 11 Pro; llama.cpp b11669 CUDA, whole model on the card
- Reported total cost
- not stated
- Power
- not stated
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| gpt-oss 20B | MXFP4 (12.1 GB GGUF), reasoning low | llama.cpp b11669 (CUDA) | 238.6 | 9,940 @ 15,000-token prompt est. TTFT @8k ≈ 0.82 s (calculated from measured prefill) | context: code prompt, 32K window📏 measuredlocally test: one user, thinking off, same code prompt on every machine; middle of three runs (idle machine, 8 quiet seconds before each run); llama.cpp b11669, 32,768-token window; 238.1-239.2 over 3 runs; 16 users at once: 624 tok/s in total |
| Gemma 4 26B-A4B | UD-IQ4_XS (13.6 GB) | llama.cpp b11669 (CUDA) | 161.4 | 7,257 @ 19,000-token prompt est. TTFT @8k ≈ 1.1 s (calculated from measured prefill) | context: code prompt, 32K window📏 measuredlocally test: one user, thinking off, same code prompt on every machine; middle of three runs (idle machine, 8 quiet seconds before each run); llama.cpp b11669, 32,768-token window; 16 users at once: 383 tok/s in total |
| Qwen3.5 9B | Q4_K_M (5.7 GB) | llama.cpp b11669 (CUDA) | 137.6 | 8,584 @ 19,000-token prompt est. TTFT @8k ≈ 0.95 s (calculated from measured prefill) | context: code prompt, 32K window📏 measuredlocally test: one user, thinking off, same code prompt on every machine; middle of three runs (idle machine, 8 quiet seconds before each run); llama.cpp b11669, 32,768-token window; 16 users at once: 576 tok/s in total |
| Gemma 4 12B | QAT Q4_0 (7.0 GB, Google's file) | llama.cpp b11669 (CUDA) | 99.7 | 5,072 @ 19,000-token prompt est. TTFT @8k ≈ 1.6 s (calculated from measured prefill) | context: code prompt, 32K window📏 measuredlocally test: one user, thinking off, same code prompt on every machine; middle of three runs (idle machine, 8 quiet seconds before each run); llama.cpp b11669, 32,768-token window; 16 users at once: 307 tok/s in total |
| Qwen3.8 27B | Q4_K_M (16.8 GB, LM Studio catalogue) + the model's own MTP (LM Studio default) | LM Studio 2.55.0 (llama.cpp CUDA) | 62.4 | 1,994 @ 19,000-token prompt est. TTFT @8k ≈ 4.1 s (calculated from measured prefill) | ⚡ spec-decode (MTP)context: code prompt, 20K window📏 measuredlocally test as above but in LM Studio with default settings, 20,480-token window, tested 11 Oct 2026; MTP kept 78% of its guesses; 45.0 tok/s on the prose prompt |
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.
measuredlocally.com (Alex, UK), tested 20 Oct 2026 (Qwen3.8 27B on 11 Oct), full parts list published. Every model sits wholly in the 24 GB card. Artificial Analysis's AgentPerf-Local coding-agent replay with Qwen3.8 27B + MTP took 7.3 min on this machine. A ~1.75-bit Ternary Bonsai 2 27B run (105.5 tok/s) is below Q4 and not listed.
- 1× RTX 4090: ≈€3,955 each (≈$4,433), new, approx. price as of 2026-10-16, lowest listing on billiger.de (ASUS TUF Gaming RTX 4090 OG OC 24 GB)
New retail prices in Germany incl. VAT; used cards are often much cheaper. Prices change daily.
https://measuredlocally.com/hardware-leaderboard.html
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.
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
[](https://getrigbase.com/builds/rtx-4090-ryzen-9-7950x-96gb-measuredlocally/)
HTML
<a href="https://getrigbase.com/builds/rtx-4090-ryzen-9-7950x-96gb-measuredlocally/" rel="noopener"><img src="https://getrigbase.com/badge/rtx-4090-ryzen-9-7950x-96gb-measuredlocally.svg" alt="Measured on RigBase" height="20"/></a>
Preview: 
· How to use
Quick answers: Can RTX 4090 run Gemma 4 12B? · Can RTX 4090 run Gemma 4 26B-A4B? · Can RTX 4090 run Qwen3.5 9B? · Can RTX 4090 run Qwen3.8 27B? · Can RTX 4090 run gpt-oss 20B?