Can RTX 5090 32GB run Qwen3.8 27B?
Measured results for Qwen3.8 27B (27.78B dense) on RTX 5090 32GB, plus what fits and the commands to run it.
Short answer
Yes. Qwen3.8 27B Q8_0 fits fully on RTX 5090 32GB at 16k context, using 30.3 of 31.0 GB usable, and generates about 53 tokens/s in our own run. As the context fills it slows down: 51.3 tok/s with 16k tokens already in context (97% of the empty-context speed) and 49.6 tok/s at 32k (94%), measured. For Q8_0, our bandwidth-only estimate was ~34 tok/s against a measured median of ~53 (36% too pessimistic).
Measured generation speed by quantization
| Quant | Generation tok/s | Range | Prompt tok/s | Runs | Origin |
|---|---|---|---|---|---|
| Q2 class (median) | 109 | 103.3–114.6 | 3621 | 2 | public runs |
| Q3 class (median) | 96.2 | 96.2–96.2 | 3944 | 1 | public runs |
| Q4 class (median) | 78.9 | 76.2–82.5 | 3881 | 3 | includes our runs |
| Q6 class (median) | 63.1 | 63.1–63.1 | 3223 | 1 | public runs |
| Q8 class (median) | 53 | 53–53.1 | 4106 | 2 | includes our runs |
| Q4_K_M (our run, depth 0) | 82.5 | 3 repeats | 4629 | 3 | own measurement |
| Q8_0 (our run, depth 0) | 53 | 3 repeats | 4378 | 3 | own measurement |
Generation = tg128 (128 generated tokens, short context). Public rows are medians of community llama.cpp results (sources); our own runs use llama-bench b11658 on a rented GPU.
Speed vs context depth measured
Our own llama-bench runs on GeForce RTX 5090 (flash attention on, 3 repeats): generation and prompt speed with the given number of tokens already in the context.
| Quant | Context depth | Generation tok/s | vs depth 0 | Prompt tok/s |
|---|---|---|---|---|
| Q4_K_M | 0 | 82.5 | 100% | 4629 |
| Q4_K_M | 4k | 81.6 | 99% | 4502 |
| Q4_K_M | 16k | 78.3 | 95% | 4031 |
| Q4_K_M | 32k | 74.5 | 90% | 3317 |
| Q4_K_M | 64k | 67.9 | 82% | 2107 |
| Q8_0 | 0 | 53 | 100% | 4378 |
| Q8_0 | 4k | 52.5 | 99% | 4294 |
| Q8_0 | 16k | 51.3 | 97% | 3887 |
| Q8_0 | 32k | 49.6 | 94% | 3232 |
Alternatives
Less memory that still fits Qwen3.8 27B
Smaller memory usually means a cheaper card or Mac (we do not list prices).
- Mac mini M6 32GB: UD-Q4_K_M, 22 GB usable, ~6 tok/s estimated
- MacBook Air M5 32GB: UD-Q4_K_M, 22 GB usable, ~6 tok/s estimated
- RTX 4090 24GB: UD-Q5_K_M, 23 GB usable, ~44 tok/s scaled from measured
Other models on RTX 5090 32GB, similar size
- Qwen3.6 27B: Q8_0, fits fully, ~52 tok/s (measured)
- Gemma 4 26B-A4B: Q8_0, fits fully, ~217 tok/s (measured)
- Qwen3-Coder 30B-A3B: Q6_K, fits fully, ~256 tok/s (scaled from measured)
- GLM-4.7-Flash (hardware page): Q6_K, fits fully, ~268 tok/s (scaled from measured)
Memory and context limits
Q8_0 weights are 29.1 GB; the 16k f16 KV cache adds 1.23 GB. Largest context that fits in 31.0 GB usable:
| KV cache type | Max context |
|---|---|
| f16 | 16k |
| q8_0 | 32k |
| q4_0 | 64k |
Native limit 256k, 1M with YaRN. Full KV cache table for Qwen3.8 27B.
Copy-paste commands
The advisor's pick for this hardware: Q8_0, 16k context.
llama-server -hf unsloth/Qwen3.8-27B-GGUF:Q8_0 -c 16384 -fa on --jinja -ngl 99
OLLAMA_CONTEXT_LENGTH=16384 ollama serve ollama run hf.co/unsloth/Qwen3.8-27B-GGUF:Q8_0
FAQ
How fast is Qwen3.8 27B on RTX 5090 32GB?
About 53 tokens/s for generation (Q8_0 on llama.cpp CUDA, 128-token generation), 4378 tokens/s for prompt processing. Speed drops as the context fills; see the depth table.
How much memory does Qwen3.8 27B need on RTX 5090 32GB?
The advisor's pick, Q8_0, is 29.1 GB of weights plus 1.23 GB of KV cache at 16k context (30.3 GB), against 31.0 GB usable.
Is the Qwen3.8 27B speed on RTX 5090 32GB measured or estimated?
Measured: our own llama-bench run on a rented GeForce RTX 5090, including runs with 4k to 64k tokens already in context.
What can I use instead of RTX 5090 32GB for Qwen3.8 27B?
With less memory, Mac mini M6 32GB (~6 tok/s), MacBook Air M5 32GB (~6 tok/s), RTX 4090 24GB (~44 tok/s) also fit it fully.
What is the maximum context for Qwen3.8 27B on RTX 5090 32GB?
With an f16 KV cache about 16k tokens, 32k with q8_0 and 64k with q4_0, bounded by the model's own limit.
Related
All models on RTX 5090 32GB · Qwen3.8 27B on all hardware · VRAM tier · all measured pages · leaderboard
Compare: RTX 4090 24GB vs RTX 5090 32GB · RTX 3090 24GB vs RTX 5090 32GB · 2x RTX 3090 (48GB total) vs RTX 5090 32GB