Can I run Gemma 4 31B on an RTX 4090?
Only with a shorter context: Gemma 4 31B needs about 23.9 GB at Q4_K_M with 32K tokens of context, which leaves only 80 MB free on the 24 GB RTX 4090, too tight to count as a fit, but Q4_K_M fits with up to 26K tokens. The smallest setup here that holds Gemma 4 31B at Q4_K_M with 32K is RTX 3090 (24 GB).
Partly Q4_K_M with 26K tokens of context
- Q4_K_M, 32K
- 23.9 GB
- RTX 4090
- 24 GB, 1,008 GB/s
- Free (tight)
- 80 MB
- Tokens/s
- 24–33 tokens/s
At Q4_K_M with 26K tokens of context it writes about 24–33 tokens/s for one request on an RTX 4090.
Best precision for Gemma 4 31B on an RTX 4090
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 | Q4_K_M | 21.9 GB | 2.14 GB | 26–36 |
| 32K | Q3_K_M | 20.2 GB | 3.80 GB | 28–39 |
| 128K | Nothing fits | — | — | — |
| 256K (full) | Nothing fits | — | — | — |
Gemma 4 31B on the RTX 4090 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 | 21.5 GB | 2.48 GB | 26–36 | 36.2 GB | −12.2 GB | — |
| 8K | 21.9 GB | 2.14 GB | 26–36 | 36.5 GB | −12.5 GB | — |
| 16K | 22.5 GB | 1.45 GB | 25–35 | 37.2 GB | −13.2 GB | — |
| 32K | 23.9 GB | 80 MB (tight) | 23–33 | 38.6 GB | −14.6 GB | — |
| 64K | 26.7 GB | −2.67 GB | — | 41.3 GB | −17.3 GB | — |
| 128K | 32.2 GB | −8.17 GB | — | 46.8 GB | −22.8 GB | — |
| 256K (full) | 43.2 GB | −19.2 GB | — | 57.8 GB | −33.8 GB | — |
Gemma 4 31B on more than one RTX 4090
| Cards | Best at 32K | Tokens/s | Q4_K_M longest | Q4_K_M tokens/s | Rent per hour |
|---|---|---|---|---|---|
| 2× (48 GB) | Q8_0 | 27–39 | 256K (full) | 41–61 | $1.06 |
| 4× (96 GB) | BF16 | 30–43 | 256K (full) | 72–113 | $2.12 |
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 20.2 GB at 32K, 3.80 GB under the RTX 4090; it fits with 0.5 GB to spare up to 70K tokens.
- Renting an RTX 4090 costs about $0.53 an hour (median on getdeploying.com, 2026-09-29): $4.4–6.2 per million tokens at 24–33 tokens/s.
- Smallest setup for Q4_K_M at 32K: RTX 3090 (24 GB).
Run Gemma 4 31B on the RTX 4090 with llama-server
llama-server -hf unsloth/gemma-4-31B-it-GGUF:Q4_K_M -c 26624 -np 1gemma-4-31B-it-Q4_K_M.gguf, 18.3 GB, from unsloth/
Questions
Can I run Gemma 4 31B on an RTX 4090?
Only with a shorter context: Gemma 4 31B needs about 23.9 GB at Q4_K_M with 32K tokens of context, which leaves only 80 MB free on the 24 GB RTX 4090, too tight to count as a fit, but Q4_K_M fits with up to 26K tokens. The smallest setup here that holds Gemma 4 31B at Q4_K_M with 32K is RTX 3090 (24 GB).
How fast is Gemma 4 31B on an RTX 4090?
At Q4_K_M with 26K tokens of context it writes about 24–33 tokens/s for one request on an RTX 4090.
What does a second RTX 4090 change for Gemma 4 31B?
Two RTX 4090 cards (48 GB in one tensor-parallel group) hold Gemma 4 31B at Q8_0 with 32K, and Q4_K_M up to 256K (full) tokens, at about 41–61 tokens/s.
Try other settings in the VRAM calculator or the speed calculator. See also Gemma 4 31B VRAM requirements, what LLMs an RTX 4090 can run and every pair, or detect your own GPU. Model data checked .