Can 8GB VRAM class (RTX 4060 8GB, 32GB RAM) run Qwen3.5 9B?
Measured results for Qwen3.5 9B (9.65B dense) on 8GB VRAM class (RTX 4060 8GB, 32GB RAM), plus what fits and the commands to run it.
Short answer
Yes. Qwen3.5 9B Q4_K_M fits fully on 8GB VRAM class (RTX 4060 8GB, 32GB RAM) at 16k context, using 6.3 of 7.0 GB usable, and generates about 45.6 tokens/s in our own run. As the context fills it slows down: 41.6 tok/s with 16k tokens already in context (91% of the empty-context speed), measured. For Q4_K_M, our bandwidth-only estimate was ~26 tok/s against a measured median of ~46 (42% too pessimistic).
Measured generation speed by quantization
| Quant | Generation tok/s | Range | Prompt tok/s | Runs | Origin |
|---|---|---|---|---|---|
| Q4 class (median) | 45.6 | 45.6–45.6 | 2283 | 1 | includes our runs |
| Q4_K_M (our run, depth 0) | 45.6 | 3 repeats | 2283 | 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 4060 (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 | 45.6 | 100% | 2283 |
| Q4_K_M | 4k | 44.5 | 98% | 2213 |
| Q4_K_M | 16k | 41.6 | 91% | 1997 |
Alternatives
Faster for Qwen3.5 9B
- RTX 5070 12GB: Q8_0, 11 GB usable, ~68 tok/s measured
- RTX 5080 16GB: Q8_0, 15 GB usable, ~92 tok/s measured
- RTX 5070 Ti 16GB: Q8_0, 15 GB usable, ~86 tok/s measured
Other models on 8GB VRAM class (RTX 4060 8GB, 32GB RAM), similar size
- Ornith 1.5 9B (hardware page): Q4_K_M, fits fully, ~44 tok/s (scaled from measured)
- Gemma 4 E4B (hardware page): Q5_K_M, fits fully, ~47 tok/s (scaled from measured)
- DeepSeek R1 Distill Qwen 7B (hardware page): Q5_K_M, fits fully, ~48 tok/s (scaled from measured)
- Qwen3.5 4B: Q8_0, fits fully, ~51 tok/s (measured)
Memory and context limits
Q4_K_M weights are 5.7 GB; the 16k f16 KV cache adds 0.59 GB. Largest context that fits in 7.0 GB usable:
| KV cache type | Max context |
|---|---|
| f16 | 32k |
| q8_0 | 64k |
| q4_0 | 128k |
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: Q4_K_M, 16k context.
llama-server -hf unsloth/Qwen3.5-9B-GGUF:Q4_K_M -c 16384 -fa on --jinja -ngl 99
OLLAMA_CONTEXT_LENGTH=16384 ollama serve ollama run hf.co/unsloth/Qwen3.5-9B-GGUF:Q4_K_M
FAQ
How fast is Qwen3.5 9B on 8GB VRAM class (RTX 4060 8GB, 32GB RAM)?
About 45.6 tokens/s for generation (Q4_K_M on llama.cpp CUDA, 128-token generation), 2283 tokens/s for prompt processing. Speed drops as the context fills; see the depth table.
How much memory does Qwen3.5 9B need on 8GB VRAM class (RTX 4060 8GB, 32GB RAM)?
The advisor's pick, Q4_K_M, is 5.7 GB of weights plus 0.59 GB of KV cache at 16k context (6.3 GB), against 7.0 GB usable.
Is the Qwen3.5 9B speed on 8GB VRAM class (RTX 4060 8GB, 32GB RAM) measured or estimated?
Measured: our own llama-bench run on a rented GeForce RTX 4060, including runs with 4k to 16k tokens already in context.
What can I use instead of 8GB VRAM class (RTX 4060 8GB, 32GB RAM) for Qwen3.5 9B?
Nothing with less memory in our list fits it fully. Faster: RTX 5070 12GB (~68 tok/s), RTX 5080 16GB (~92 tok/s), RTX 5070 Ti 16GB (~86 tok/s).
What is the maximum context for Qwen3.5 9B on 8GB VRAM class (RTX 4060 8GB, 32GB RAM)?
With an f16 KV cache about 32k tokens, 64k with q8_0 and 128k with q4_0, bounded by the model's own limit.
Related
All models on 8GB VRAM class (RTX 4060 8GB, 32GB RAM) · Qwen3.5 9B on all hardware · VRAM tier · all measured pages · leaderboard
Compare: RTX 3060 12GB (32GB RAM) vs 8GB VRAM class (RTX 4060 8GB, 32GB RAM)