Can I run Qwen3.8 27B on an RTX 4090?
Yes: Qwen3.8 27B needs about 19.9 GB at Q4_K_M with 32K tokens of context, which fits the 24 GB RTX 4090 with 4.08 GB to spare. At 32K the RTX 4090 holds up to Q5_K_M (22.9 GB), and Q4_K_M runs up to 83K tokens.
Yes Q4_K_M with 32K tokens of context
- Q4_K_M, 32K
- 19.9 GB
- RTX 4090
- 24 GB, 1,008 GB/s
- To spare
- 4.08 GB
- Tokens/s
- 28–39 tokens/s
At Q4_K_M with 32K tokens of context it writes about 28–39 tokens/s for one request on an RTX 4090.
Best precision for Qwen3.8 27B on an RTX 4090
The most precise setting that leaves at least 0.5 GB free; one that fits with less is marked tight.
| Context | Best fit | Memory | Free | Tokens/s |
|---|---|---|---|---|
| 8K | Q5_K_M | 21.2 GB | 2.78 GB | 26–37 |
| 32K | Q5_K_M | 22.9 GB | 1.13 GB | 24–34 |
| 128K | Q3_K_M | 23.2 GB | 809 MB | 24–34 |
| 256K (full) | Nothing fits | — | — | — |
Qwen3.8 27B on the RTX 4090 as the context fills
One request, FP16 KV cache, 0.5 GB plus 10% overhead; a minus sign is memory missing, and tight is less than 0.5 GB free.
| Context | Q4_K_M | Free | Tokens/s | Q8_0 | Free | Tokens/s |
|---|---|---|---|---|---|---|
| 4K | 18.0 GB | 6.01 GB | 31–43 | 31.0 GB | −7.01 GB | — |
| 8K | 18.3 GB | 5.73 GB | 31–43 | 31.3 GB | −7.29 GB | — |
| 16K | 18.8 GB | 5.18 GB | 30–41 | 31.8 GB | −7.84 GB | — |
| 32K | 19.9 GB | 4.08 GB | 28–39 | 32.9 GB | −8.94 GB | — |
| 64K | 22.1 GB | 1.88 GB | 25–35 | 35.1 GB | −11.1 GB | — |
| 128K | 26.5 GB | −2.52 GB | — | 39.5 GB | −15.5 GB | — |
| 256K (full) | 35.3 GB | −11.3 GB | — | 48.3 GB | −24.3 GB | — |
Qwen3.8 27B on more than one RTX 4090
| Cards | Best at 32K | Tokens/s | Q4_K_M longest | Q4_K_M tokens/s | Rent per hour |
|---|---|---|---|---|---|
| 2× (48 GB) | Q8_0 | 31–45 | 256K (full) | 49–72 | $1.06 |
| 4× (96 GB) | BF16 | 34–49 | 256K (full) | 83–132 | $2.12 |
Tensor parallel with the combined bandwidth, as in vLLM; llama.cpp splits layers by default and runs at about one card's speed.
Other options
- The next smaller setting, Q3_K_M, takes 16.6 GB at 32K, 7.39 GB under the RTX 4090; it fits with 0.5 GB to spare up to 131K tokens.
- Renting an RTX 4090 costs about $0.53 an hour (median on getdeploying.com, 2026-09-29): $3.8–5.3 per million tokens at 28–39 tokens/s.
- Smallest setup for Q4_K_M at 32K: RTX 3090 (24 GB).
Run Qwen3.8 27B on the RTX 4090 with llama-server
llama-server -hf ggml-org/Qwen3.8-27B-GGUF:Q4_K_M -c 52224Qwen3.8-27B-Q4_K_M.gguf, 19.0 GB, from ggml-org/
Questions
Can I run Qwen3.8 27B on an RTX 4090?
Yes: Qwen3.8 27B needs about 19.9 GB at Q4_K_M with 32K tokens of context, which fits the 24 GB RTX 4090 with 4.08 GB to spare. At 32K the RTX 4090 holds up to Q5_K_M (22.9 GB), and Q4_K_M runs up to 83K tokens.
How fast is Qwen3.8 27B on an RTX 4090?
At Q4_K_M with 32K tokens of context it writes about 28–39 tokens/s for one request on an RTX 4090.
What does a second RTX 4090 change for Qwen3.8 27B?
Two RTX 4090 cards (48 GB in one tensor-parallel group) hold Qwen3.8 27B at Q8_0 with 32K, and Q4_K_M up to 256K (full) tokens, at about 49–72 tokens/s.
Try other settings in the VRAM calculator or the speed calculator. See also Qwen3.8 27B VRAM requirements, what LLMs an RTX 4090 can run and every pair, or detect your own GPU. Model data checked .