Can RTX 5090 32GB run Gemma 4 31B?
Measured results for Gemma 4 31B (31.27B dense) on RTX 5090 32GB, plus what fits and the commands to run it.
Short answer
Yes. Gemma 4 31B Q4_K_M fits fully on RTX 5090 32GB at 16k context, using 20.5 of 31.0 GB usable, and generates about 71.3 tokens/s in our own run. As the context fills it slows down: 64.6 tok/s with 16k tokens already in context (91% of the empty-context speed), measured. For Q6_K, our bandwidth-only estimate was ~39 tok/s against a measured median of ~55 (29% too pessimistic).
Measured generation speed by quantization
| Quant | Generation tok/s | Range | Prompt tok/s | Runs | Origin |
|---|---|---|---|---|---|
| Q4 class (median) | 73.8 | 71.3–76.4 | 4048 | 2 | includes our runs |
| Q6 class (median) | 55.4 | 55.4–55.4 | 3165 | 1 | public runs |
| Q4_K_M (our run, depth 0) | 71.3 | 3 repeats | 3841 | 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 5090 (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 | 71.3 | 100% | 3841 |
| Q4_K_M | 4k | 67.2 | 94% | 3607 |
| Q4_K_M | 16k | 64.6 | 91% | 2998 |
Alternatives
Less memory that still fits Gemma 4 31B
Smaller memory usually means a cheaper card or Mac (we do not list prices).
- Mac mini M6 32GB: Q4_K_M, 22 GB usable, ~6 tok/s estimated
- MacBook Air M5 32GB: Q4_K_M, 22 GB usable, ~5 tok/s estimated
- RTX 4090 24GB: Q4_K_M, 23 GB usable, ~44 tok/s measured
Other models on RTX 5090 32GB, similar size
- GLM-4.7-Flash (hardware page): Q6_K, fits fully, ~268 tok/s (scaled from measured)
- Nemotron 3 Nano 30B-A3B (hardware page): Q5_K_M, fits fully, ~342 tok/s (scaled from measured)
- Nemotron 3.5 Lightning 30B-A3B (hardware page): UD-Q5_K_M, fits fully, ~349 tok/s (measured)
- Qwen3-Coder 30B-A3B: Q6_K, fits fully, ~256 tok/s (scaled from measured)
Memory and context limits
Q6_K weights are 25.2 GB; the 16k f16 KV cache adds 2.18 GB. Largest context that fits in 31.0 GB usable:
| KV cache type | Max context |
|---|---|
| f16 | 32k |
| q8_0 | 64k |
| q4_0 | 128k |
Native limit 256k. Full KV cache table for Gemma 4 31B.
Copy-paste commands
The advisor's pick for this hardware: Q6_K, 16k context.
llama-server -hf unsloth/gemma-4-31B-it-GGUF:Q6_K -c 16384 -fa on --jinja -ngl 99
OLLAMA_CONTEXT_LENGTH=16384 ollama serve ollama run hf.co/unsloth/gemma-4-31B-it-GGUF:Q6_K
FAQ
How fast is Gemma 4 31B on RTX 5090 32GB?
About 71.3 tokens/s for generation (Q4_K_M on llama.cpp CUDA, 128-token generation), 3841 tokens/s for prompt processing. Speed drops as the context fills; see the depth table.
How much memory does Gemma 4 31B need on RTX 5090 32GB?
The advisor's pick, Q6_K, is 25.2 GB of weights plus 2.18 GB of KV cache at 16k context (27.4 GB), against 31.0 GB usable. The measured Q4_K_M run needs 20.5 GB.
Is the Gemma 4 31B speed on RTX 5090 32GB measured or estimated?
Measured: our own llama-bench run on a rented GeForce RTX 5090, including runs with 4k to 16k tokens already in context.
What can I use instead of RTX 5090 32GB for Gemma 4 31B?
With less memory, Mac mini M6 32GB (~6 tok/s), MacBook Air M5 32GB (~5 tok/s), RTX 4090 24GB (~44 tok/s) also fit it fully.
What is the maximum context for Gemma 4 31B on RTX 5090 32GB?
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 RTX 5090 32GB · Gemma 4 31B on all hardware · VRAM tier · all measured pages · leaderboard
Compare: RTX 4090 24GB vs RTX 5090 32GB · RTX 3090 24GB vs RTX 5090 32GB · 2x RTX 3090 (48GB total) vs RTX 5090 32GB