Can I run Gemma 4 26B-A4B on an RTX 4070 12GB?
With --n-cpu-moe 12: the UD-Q4_K_M GGUF keeps 11.6 GB on the RTX 4070 and 6.05 GB in system RAM, with the experts of 12 of its 30 layers moved there. Whole, Gemma 4 26B-A4B needs about 17.5 GB at Q4_K_M with 32K tokens of context, 5.50 GB more than the RTX 4070 12GB holds. The smallest setup here that holds Gemma 4 26B-A4B at Q4_K_M with 32K is RTX 3090 (24 GB).
With offload the UD-Q4_K_M GGUF with --n-cpu-moe 12 and 32K tokens of context
- On the card
- 11.6 GB
- RTX 4070
- 12 GB, 504 GB/s
- In RAM
- 6.05 GB
- Tokens/s
- 27–46 tokens/s
With the UD-Q4_K_M GGUF, --n-cpu-moe 12 and 32K tokens of context it writes about 27–46 tokens/s for one request on an RTX 4070 12GB and dual-channel DDR5-5600.
Best precision for Gemma 4 26B-A4B on an RTX 4070 12GB
The most precise setting that leaves at least 0.5 GB free; one that fits with less is marked tight.
| Context | Best fit | Memory | Free | Tokens/s |
|---|---|---|---|---|
| 8K | Only IQ3_XXS, tight | 11.9 GB | 103 MB (tight) | 64–112 |
| 32K | Nothing fits | — | — | — |
| 128K | Nothing fits | — | — | — |
| 256K (full) | Nothing fits | — | — | — |
Gemma 4 26B-A4B on the RTX 4070 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 | 16.9 GB | −4.90 GB | — | 29.0 GB | −17.0 GB | — |
| 8K | 17.0 GB | −4.99 GB | — | 29.1 GB | −17.1 GB | — |
| 16K | 17.2 GB | −5.16 GB | — | 29.3 GB | −17.3 GB | — |
| 32K | 17.5 GB | −5.50 GB | — | 29.6 GB | −17.6 GB | — |
| 64K | 18.2 GB | −6.19 GB | — | 30.3 GB | −18.3 GB | — |
| 128K | 19.6 GB | −7.57 GB | — | 31.7 GB | −19.7 GB | — |
| 256K (full) | 22.3 GB | −10.3 GB | — | 34.4 GB | −22.4 GB | — |
--n-cpu-moe for Gemma 4 26B-A4B on the RTX 4070
With --n-cpu-moe 12 the UD-Q4_K_M GGUF keeps 11.6 GB on the RTX 4070 and 6.05 GB in RAM, about 27–46 tokens/s with DDR5-5600. Measured UD-Q4_K_M file, 1 GB of buffers; tokens/s by system RAM speed.
Gemma 4 26B-A4B on more than one RTX 4070
| Cards | Best at 32K | Tokens/s | Q4_K_M longest | Q4_K_M tokens/s |
|---|---|---|---|---|
| 2× (24 GB) | Q5_K_M | 62–113 | 256K (full) | 68–124 |
| 4× (48 GB) | Q8_0 | 82–154 | 256K (full) | 110–214 |
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 14.4 GB at 32K, 2.43 GB over the RTX 4070; 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 Gemma 4 26B-A4B on the RTX 4070 with llama-server
llama-server -hf unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q4_K_M -c 32768 --n-cpu-moe 12 -np 1gemma-4-26B-A4B-it-UD-Q4_K_M.gguf, 16.9 GB, from unsloth/
Questions
Can I run Gemma 4 26B-A4B on an RTX 4070 12GB?
With --n-cpu-moe 12: the UD-Q4_K_M GGUF keeps 11.6 GB on the RTX 4070 and 6.05 GB in system RAM, with the experts of 12 of its 30 layers moved there. Whole, Gemma 4 26B-A4B needs about 17.5 GB at Q4_K_M with 32K tokens of context, 5.50 GB more than the RTX 4070 12GB holds. The smallest setup here that holds Gemma 4 26B-A4B at Q4_K_M with 32K is RTX 3090 (24 GB).
How fast is Gemma 4 26B-A4B on an RTX 4070 12GB?
With the UD-Q4_K_M GGUF, --n-cpu-moe 12 and 32K tokens of context it writes about 27–46 tokens/s for one request on an RTX 4070 12GB and dual-channel DDR5-5600.
Does Gemma 4 26B-A4B need --n-cpu-moe on an RTX 4070 12GB?
With --n-cpu-moe 12 the UD-Q4_K_M GGUF keeps 11.6 GB on the RTX 4070 and 6.05 GB in RAM, about 27–46 tokens/s with DDR5-5600.
What does a second RTX 4070 12GB change for Gemma 4 26B-A4B?
Two RTX 4070 12GB cards (24 GB in one tensor-parallel group) hold Gemma 4 26B-A4B at Q5_K_M with 32K, and Q4_K_M up to 256K (full) tokens, at about 68–124 tokens/s.
Try other settings in the VRAM calculator, the speed calculator or the MoE offload planner. See also Gemma 4 26B-A4B VRAM requirements, what LLMs an RTX 4070 12GB can run and every pair, or detect your own GPU. Model data checked .