estimate
Can I run Qwen3.8-27B on a Dual RTX 4090?
Yes, fully in VRAM (estimate)
Can I run Qwen3.8-27B on a Dual RTX 4090? Yes, fully in VRAM (estimate). About 17.8 GB at Q4_K_M (estimate).
VRAM per quant for Qwen3.8-27B
Updated . The later of the hardware price date and the formula date.
| Quant | Context | Weights | KV | Overhead | Total | Usable | Fit | Est. tok/s |
|---|---|---|---|---|---|---|---|---|
| Q4 Q4_K_M | 8,192 tokens | 16.5 GB measured GGUF file size | 0.5 GB estimate | 0.8 GB estimate | 17.8 GB estimate | 46.8 GB estimate | full | 50 tok/s estimate |
| Q5 Q5 (~0.71 bytes/param) | 8,192 tokens | 19.7 GB estimate | 0.5 GB estimate | 0.8 GB estimate | 21.1 GB estimate | 46.8 GB estimate | full | 42 tok/s estimate |
| Q8 Q8 | 8,192 tokens | 29.7 GB estimate | 0.5 GB estimate | 0.8 GB estimate | 31.1 GB estimate | 46.8 GB estimate | full | 28 tok/s estimate |
| FP16 FP16 | 8,192 tokens | 55.6 GB estimate | 0.5 GB estimate | 0.8 GB estimate | 56.9 GB estimate | 46.8 GB estimate | offload | 3.3 tok/s estimate |
| Q4, longer context | 32,768 tokens | 16.5 GB measured GGUF file size | 2.1 GB estimate | 0.8 GB estimate | 19.4 GB estimate | 46.8 GB estimate | full | 46 tok/s estimate |
Formula estimate, not a benchmark. Method →
community measurement, not ours
Canonical single-GPU figures on that results page: mean of five runs, one RTX 3090, the other replica stopped, the ~1K-token prompt column, 512-token greedy completions, reasoning disabled, 250 W power cap. The page says every rung is the same 4-bit checkpoint. llama.cpp out of the box 27.6 tokens per second; llama.cpp tuned (FA + q8 KV + MTP) 42.6; stock vLLM + MTP k=4 85.6; patched vLLM + MTP k=4 167.8; patched vLLM + DFlash2 k=7 212.1. The page dates the single-replica correction 31 August. results.html was last committed on 18 September 2026, which is the published date stored here. Not a localLLMrig measurement. Calculator estimates keep benchmarkSourceUrl null.
| Hardware | Tokens per second |
|---|---|
| RTX 3090, llama.cpp out of the box | 27.6 tok/s community measurement, not ours |
| RTX 3090, llama.cpp tuned (FA + q8 KV + MTP) | 42.6 tok/s community measurement, not ours |
| RTX 3090, stock vLLM + MTP k=4 | 85.6 tok/s community measurement, not ours |
| RTX 3090, patched vLLM + MTP k=4 | 167.8 tok/s community measurement, not ours |
| RTX 3090, patched vLLM + DFlash2 k=7 | 212.1 tok/s community measurement, not ours |
Qwen3.8-27B on RTX 3090, published 2026-09-18
This measurement stays separate from the formula estimate above.
Cheapest catalog option that fits at Q4
GeForce RTX 3090 24GB, used (median of sold listings), €1,224, dated 2026-10-16. No buy link.
This hardware
$8,206€7,330
price as of 16 Oct 2026
128 GB DDR4 (estimate): $2,754€2,460. From the 16 Oct 2026 fact-check.
rest of system (estimate): $560€500. The total is the GPU price plus the RAM line plus this estimate. It is not a checked system price.
price unverifiedGPU line is 2× the used median €2,185 (€4,370), labelled used, from pcprice.watch (https://www.pcprice.watch/gpu-buying-guide/rtx-4090-price-used-and-specs). 128 GB DDR4 (estimate) €2,460, from the 16 Oct 2026 fact-check. Rest of system (estimate) €500. Other parts from €500 in that fact-check. Not a shop listing. DDR5 was not separately priced, so this total uses the sourced 128 GB DDR4 line. The total is the GPU price plus the RAM line plus that rest estimate. It is not a checked system price.
Embed this
Qwen3.8-27B on Dual RTX 4090
[](https://localllmrig.com/can-i-run/qwen3.8-27b/dual-rtx-4090?utm_source=badge)
Questions
- Can I run Qwen3.8-27B on a Dual RTX 4090?
- Yes, fully in VRAM (estimate).
- How much memory does Qwen3.8-27B need at Q4?
- About 17.8 GB at 8,192 tokens (estimate): 16.5 GB weights (measured GGUF file size), 0.5 GB KV cache, 0.8 GB overhead.
- How fast is Qwen3.8-27B on Dual RTX 4090 (estimate)?
- About 50 tok/s at Q4 (estimate, not a benchmark).
Rig Report waitlist
A written hardware plan for one model and budget. Planned price, not for sale yet.
Sign up with consent and download GPU Buying Cheat Sheet (EU, Oct 2026). The file is not emailed.
Read next: Building a local LLM rig in the RAMageddon.