Can RTX 5070 12GB run Qwen3-VL 8B?
Measured results for Qwen3-VL 8B (8.77B dense) on RTX 5070 12GB, plus what fits and the commands to run it.
Short answer
Yes. Qwen3-VL 8B Q4_K_M fits fully on RTX 5070 12GB at 16k context, using 7.4 of 11.0 GB usable, and generates about 111.9 tokens/s in our own run. As the context fills it slows down: 77.1 tok/s with 16k tokens already in context (69% of the empty-context speed), measured. For Q6_K, our bandwidth-only estimate was ~55 tok/s against a measured median of ~94 (41% too pessimistic).
Measured generation speed by quantization
| Quant | Generation tok/s | Range | Prompt tok/s | Runs | Origin |
|---|---|---|---|---|---|
| Q4 class (median) | 111.9 | 111.9–111.9 | 4912 | 1 | includes our runs |
| Q8 class (median) | 72.4 | 72.4–72.4 | 5041 | 1 | includes our runs |
| Q4_K_M (our run, depth 0) | 111.9 | 3 repeats | 4912 | 3 | own measurement |
| Q8_0 (our run, depth 0) | 72.4 | 3 repeats | 5041 | 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 5070 (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 | 111.9 | 100% | 4912 |
| Q4_K_M | 4k | 100.6 | 90% | 4254 |
| Q4_K_M | 16k | 77.1 | 69% | 2916 |
| Q8_0 | 0 | 72.4 | 100% | 5041 |
| Q8_0 | 4k | 67.5 | 93% | 4358 |
| Q8_0 | 16k | 56.3 | 78% | 2971 |
Alternatives
Faster for Qwen3-VL 8B
- RTX 5090 32GB: Q8_0, 31 GB usable, ~167 tok/s measured
- RTX PRO 6000 96GB: Q8_0, 95 GB usable, ~163 tok/s measured
Other models on RTX 5070 12GB, similar size
- Gemma 4 E4B (hardware page): Q8_0, fits fully, ~77 tok/s (scaled from measured)
- Qwen3.5 9B: Q8_0, fits fully, ~68 tok/s (measured)
- Ornith 1.5 9B (hardware page): Q8_0, fits fully, ~67 tok/s (scaled from measured)
- DeepSeek R1 Distill Qwen 7B (hardware page): Q8_0, fits fully, ~78 tok/s (scaled from measured)
Memory and context limits
Q6_K weights are 6.7 GB; the 16k f16 KV cache adds 2.42 GB. Largest context that fits in 11.0 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-VL 8B.
Copy-paste commands
The advisor's pick for this hardware: Q6_K, 16k context.
llama-server -hf unsloth/Qwen3-VL-8B-Instruct-GGUF:Q6_K -c 16384 -fa on --jinja -ngl 99
OLLAMA_CONTEXT_LENGTH=16384 ollama serve ollama run hf.co/unsloth/Qwen3-VL-8B-Instruct-GGUF:Q6_K
FAQ
How fast is Qwen3-VL 8B on RTX 5070 12GB?
About 111.9 tokens/s for generation (Q4_K_M on llama.cpp CUDA, 128-token generation), 4912 tokens/s for prompt processing. Speed drops as the context fills; see the depth table.
How much memory does Qwen3-VL 8B need on RTX 5070 12GB?
The advisor's pick, Q6_K, is 6.7 GB of weights plus 2.42 GB of KV cache at 16k context (9.1 GB), against 11.0 GB usable. The measured Q4_K_M run needs 7.4 GB.
Is the Qwen3-VL 8B speed on RTX 5070 12GB measured or estimated?
Measured: our own llama-bench run on a rented GeForce RTX 5070, including runs with 4k to 16k tokens already in context.
What can I use instead of RTX 5070 12GB for Qwen3-VL 8B?
Nothing with less memory in our list fits it fully. Faster: RTX 5090 32GB (~167 tok/s), RTX PRO 6000 96GB (~163 tok/s).
What is the maximum context for Qwen3-VL 8B on RTX 5070 12GB?
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 5070 12GB · Qwen3-VL 8B on all hardware · VRAM tier · all measured pages · leaderboard
Compare: RTX 4070 12GB vs RTX 5070 12GB