Can RTX 5090 32GB run Qwen3.5 9B?
Measured results for Qwen3.5 9B (9.65B dense) on RTX 5090 32GB, plus what fits and the commands to run it.
Short answer
Yes. Qwen3.5 9B Q8_0 fits fully on RTX 5090 32GB at 16k context, using 10.1 of 31.0 GB usable, and generates about 156 tokens/s in our own run. As the context fills it slows down: 147.8 tok/s with 16k tokens already in context (95% of the empty-context speed), measured. For Q8_0, our bandwidth-only estimate was ~103 tok/s against a measured median of ~156 (34% too pessimistic).
Measured generation speed by quantization
| Quant | Generation tok/s | Range | Prompt tok/s | Runs | Origin |
|---|---|---|---|---|---|
| Q4 class (median) | 220.5 | 220.5–220.5 | 11721 | 1 | includes our runs |
| Q8 class (median) | 156 | 156–156 | 11960 | 1 | includes our runs |
| Q4_K_M (our run, depth 0) | 220.5 | 3 repeats | 11721 | 3 | own measurement |
| Q8_0 (our run, depth 0) | 156 | 3 repeats | 11960 | 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 | 220.5 | 100% | 11721 |
| Q4_K_M | 4k | 213.8 | 97% | 11596 |
| Q4_K_M | 16k | 205.2 | 93% | 10365 |
| Q8_0 | 0 | 156 | 100% | 11960 |
| Q8_0 | 4k | 152.1 | 98% | 11881 |
| Q8_0 | 16k | 147.8 | 95% | 10618 |
Alternatives
Less memory that still fits Qwen3.5 9B
Smaller memory usually means a cheaper card or Mac (we do not list prices).
- 8GB VRAM class (RTX 4060 8GB, 32GB RAM): Q4_K_M, 7 GB usable, ~46 tok/s measured
- Mac mini / iMac M4 16GB: Q4_K_M, 11 GB usable, ~14 tok/s scaled from measured
- RTX 5070 12GB: Q8_0, 11 GB usable, ~68 tok/s measured
Other models on RTX 5090 32GB, similar size
- Ornith 1.5 9B (hardware page): Q8_0, fits fully, ~145 tok/s (scaled from measured)
- Qwen3-VL 8B: Q8_0, fits fully, ~167 tok/s (measured)
- Gemma 4 E4B (hardware page): Q8_0, fits fully, ~177 tok/s (scaled from measured)
- Gemma 4 12B: Q8_0, fits fully, ~104 tok/s (measured)
Memory and context limits
Q8_0 weights are 9.5 GB; the 16k f16 KV cache adds 0.59 GB. Largest context that fits in 31.0 GB usable:
| KV cache type | Max context |
|---|---|
| f16 | 512k |
| q8_0 | 1M |
| q4_0 | 1M |
Native limit 256k, 1.01M with YaRN. Full KV cache table for Qwen3.5 9B.
Copy-paste commands
The advisor's pick for this hardware: Q8_0, 16k context.
llama-server -hf unsloth/Qwen3.5-9B-GGUF:Q8_0 -c 16384 -fa on --jinja -ngl 99
OLLAMA_CONTEXT_LENGTH=16384 ollama serve ollama run hf.co/unsloth/Qwen3.5-9B-GGUF:Q8_0
FAQ
How fast is Qwen3.5 9B on RTX 5090 32GB?
About 156 tokens/s for generation (Q8_0 on llama.cpp CUDA, 128-token generation), 11960 tokens/s for prompt processing. Speed drops as the context fills; see the depth table.
How much memory does Qwen3.5 9B need on RTX 5090 32GB?
The advisor's pick, Q8_0, is 9.5 GB of weights plus 0.59 GB of KV cache at 16k context (10.1 GB), against 31.0 GB usable.
Is the Qwen3.5 9B 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 16k tokens already in context.
What can I use instead of RTX 5090 32GB for Qwen3.5 9B?
With less memory, 8GB VRAM class (RTX 4060 8GB, 32GB RAM) (~46 tok/s), Mac mini / iMac M4 16GB (~14 tok/s), RTX 5070 12GB (~68 tok/s) also fit it fully.
What is the maximum context for Qwen3.5 9B on RTX 5090 32GB?
With an f16 KV cache about 512k tokens, 1M with q8_0 and 1M with q4_0, bounded by the model's own limit.
Related
All models on RTX 5090 32GB · Qwen3.5 9B 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