Gemma · 31.3B parameters
Gemma 4 31B: GPU memory requirements and which GPU to rent
Google's Gemma 4 31B, released in 2026: 60 layers, 50 of them on a 1,024-token sliding window and 10 global layers with large heads.
- Weights at FP831 GB63 GB at 16-bit · 17 GB at 4-bit
- FP8, 8k context, 1 sequence36 GBweights + KV cache + runtime
- Suggested here1× L40S$429 a month · dedicated
Inference
How much GPU memory Gemma 4 31B needs to serve
Total memory for one sequence at each context length, and for 16 concurrent sequences at 8k — weights, KV cache at 16-bit and runtime overhead included.
| Weights precision | Weights | 4k context | 32k context | 128k context | 8k × 16 sequences | Suggested here, 8k |
|---|---|---|---|---|---|---|
| 16-bit (BF16 / FP16) | 63 GB | 68 GB | 70 GB | 78 GB | 91 GB | 1× A100 · $649/mo |
| 8-bit (FP8 / INT8) | 31 GB | 35 GB | 38 GB | 46 GB | 58 GB | 1× L40S · $429/mo |
| 4-bit (GPTQ, AWQ, Q4) | 17 GB | 21 GB | 23 GB | 31 GB | 44 GB | 1× L4 · $169/mo |
GB are 109 bytes. Context lengths are shown for scale where the model supports them; an FP8 KV cache halves the cache part of each figure. Fifty of its sixty layers use a 1,024-token sliding window.
Hardware
GPU servers that run Gemma 4 31B
For FP8 inference with an 8k context and one sequence: the configurations we would suggest, in order — one large card ahead of several small ones.
Fine-tuning Gemma 4 31B
Memory for a 2k-token sequence and a micro-batch of 4, with gradient checkpointing. Full fine-tuning keeps gradients and Adam states for every parameter; LoRA and QLoRA train small adapters on a frozen base.
| Method | Memory | Suggested here |
|---|---|---|
| QLoRA, 4-bit base | 32 GB | 1× L40S · $429/mo |
| LoRA, 16-bit base | 79 GB | 1× RTX PRO 6000 · $749/mo |
| Full fine-tune, mixed precision | 515 GB | 4× H200 · $6,760/mo |
Architecture at a glance
Fifty of its sixty layers use a 1,024-token sliding window. Figures from the model's published configuration; method on the calculator page.
Gemma 4 31B hardware: questions answered
How much VRAM does Gemma 4 31B need?
For the weights alone: about 63 GB at 16-bit, 31 GB at FP8 and 17 GB at 4-bit. Serving it at FP8 with an 8k-token context and one sequence takes about 36 GB in total once the KV cache and the runtime are counted; long contexts and many concurrent users add the KV cache on top — about 0.92 GB per 1,000 tokens at 16-bit.
Can Gemma 4 31B run on a single GPU?
Yes, at FP8 with a moderate context: one NVIDIA L40S (48 GB GDDR6) holds it, for $429 a month here. Longer contexts or more concurrent users may need a larger card or a second GPU.
What is the cheapest GPU setup to run Gemma 4 31B?
For FP8 inference with an 8k context and one sequence, the configuration we would suggest is 1× L40S · $429/mo; at 4-bit, 1× L4 · $169/mo. Suggestions favour one large card over several small ones: splitting a model across PCIe cards costs speed.
How much GPU memory do I need to fine-tune Gemma 4 31B?
With a 2k-token sequence and a micro-batch of 4: about 32 GB for QLoRA on a 4-bit base, 79 GB for LoRA on a 16-bit base, and 515 GB for a full fine-tune with Adam in mixed precision. Gradient checkpointing is assumed; sharding the optimizer across GPUs spreads the full fine-tune over several cards.
Run Gemma 4 31B on 1× L40S for $429 a month
Dedicated, single-tenant and live in under 10 minutes — with vLLM, the Hugging Face stack or Ollama preinstalled if you want. Paid in BTC, ETH or USDT, with no KYC.