Can I run Qwen3.6 35B-A3B on an RTX 3060 12GB?
With --n-cpu-moe 22: the UD-Q4_K_M GGUF keeps 11.8 GB on the RTX 3060 and 10.5 GB in system RAM, with the experts of 22 of its 40 layers moved there. Whole, Qwen3.6 35B-A3B needs about 23.5 GB at Q4_K_M with 32K tokens of context, 11.5 GB more than the RTX 3060 12GB holds. The smallest setup here that holds Qwen3.6 35B-A3B at Q4_K_M with 32K is RTX 3090 (24 GB).
With offload the UD-Q4_K_M GGUF with --n-cpu-moe 22 and 32K tokens of context
- On the card
- 11.8 GB
- RTX 3060
- 12 GB, 360 GB/s
- In RAM
- 10.5 GB
- Tokens/s
- 24–41 tokens/s
With the UD-Q4_K_M GGUF, --n-cpu-moe 22 and 32K tokens of context it writes about 24–41 tokens/s for one request on an RTX 3060 12GB and dual-channel DDR5-5600.
Qwen3.6 35B-A3B on the RTX 3060 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 | 22.9 GB | −10.9 GB | — | 39.7 GB | −27.7 GB | — |
| 8K | 23.0 GB | −11.0 GB | — | 39.8 GB | −27.8 GB | — |
| 16K | 23.1 GB | −11.1 GB | — | 40.0 GB | −28.0 GB | — |
| 32K | 23.5 GB | −11.5 GB | — | 40.3 GB | −28.3 GB | — |
| 64K | 24.2 GB | −12.2 GB | — | 41.0 GB | −29.0 GB | — |
| 128K | 25.5 GB | −13.5 GB | — | 42.4 GB | −30.4 GB | — |
| 256K (full) | 28.3 GB | −16.3 GB | — | 45.1 GB | −33.1 GB | — |
--n-cpu-moe for Qwen3.6 35B-A3B on the RTX 3060
With --n-cpu-moe 22 the UD-Q4_K_M GGUF keeps 11.8 GB on the RTX 3060 and 10.5 GB in RAM, about 24–41 tokens/s with DDR5-5600. Measured UD-Q4_K_M file, 1 GB of buffers; tokens/s by system RAM speed.
Qwen3.6 35B-A3B on more than one RTX 3060
| Cards | Best at 32K | Tokens/s | Q4_K_M longest | Q4_K_M tokens/s | Rent per hour |
|---|---|---|---|---|---|
| 2× (24 GB) | Q3_K_M | 75–138 | 9K | — | $0.16 |
| 4× (48 GB) | Q8_0 | 80–150 | 256K (full) | 108–210 | $0.32 |
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 19.2 GB at 32K, 7.19 GB over the RTX 3060; it does not fit with 0.5 GB to spare even with 1K tokens.
- Smallest setup for Q4_K_M at 32K: RTX 3090 (24 GB).
Run Qwen3.6 35B-A3B on the RTX 3060 with llama-server
llama-server -hf unsloth/Qwen3.6-35B-A3B-GGUF:UD-Q4_K_M -c 32768 --n-cpu-moe 22Qwen3.6-35B-A3B-UD-Q4_K_M.gguf, 22.1 GB, from unsloth/
Questions
Can I run Qwen3.6 35B-A3B on an RTX 3060 12GB?
With --n-cpu-moe 22: the UD-Q4_K_M GGUF keeps 11.8 GB on the RTX 3060 and 10.5 GB in system RAM, with the experts of 22 of its 40 layers moved there. Whole, Qwen3.6 35B-A3B needs about 23.5 GB at Q4_K_M with 32K tokens of context, 11.5 GB more than the RTX 3060 12GB holds. The smallest setup here that holds Qwen3.6 35B-A3B at Q4_K_M with 32K is RTX 3090 (24 GB).
How fast is Qwen3.6 35B-A3B on an RTX 3060 12GB?
With the UD-Q4_K_M GGUF, --n-cpu-moe 22 and 32K tokens of context it writes about 24–41 tokens/s for one request on an RTX 3060 12GB and dual-channel DDR5-5600.
Does Qwen3.6 35B-A3B need --n-cpu-moe on an RTX 3060 12GB?
With --n-cpu-moe 22 the UD-Q4_K_M GGUF keeps 11.8 GB on the RTX 3060 and 10.5 GB in RAM, about 24–41 tokens/s with DDR5-5600.
What does a second RTX 3060 12GB change for Qwen3.6 35B-A3B?
Two RTX 3060 12GB cards (24 GB in one tensor-parallel group) hold Qwen3.6 35B-A3B at Q3_K_M with 32K, and Q4_K_M up to 9K tokens.
Try other settings in the VRAM calculator, the speed calculator or the MoE offload planner. See also Qwen3.6 35B-A3B VRAM requirements, what LLMs an RTX 3060 12GB can run and every pair, or detect your own GPU. Model data checked .