Hardware guide · October 2026
What GPU do you need to run Gemma 4 26B A4B?
Gemma 4 26B A4B is a 26B (3.8B active) mixture-of-experts model. Here is exactly how much memory it needs and which GPUs and Macs run it well.
The short answer
To run Gemma 4 26B A4B at Q4_K_M with 8K tokens of context you need about 17GB of VRAM (15.8GB of weights plus KV cache and overhead). The cheapest way in is a 16GB card like the RTX 4060 Ti 16GB (~$380 used). For the best speed per dollar, get the RTX 3090 (~110 tokens/sec). On a Mac or other unified-memory system, the Ryzen AI Max+ 395 runs it at ~26 tokens/sec.
- VRAM at Q4_K_M
- ~17GB
- GPUs that run it
- 31
- Parameters
- 26B
- CPU only (est.)
- ~18 tok/s
Our recommendations
The best GPUs for Gemma 4 26B A4B
Minimum
RTX 4060 Ti 16GB16GB · ~$380 used
- Quant
- Q3_K_M
- Speed (est.)
- ~63 tok/s
The smallest card that runs it at Q3_K_M with 8K context.
Best value
RTX 309024GB · ~$1,050 used
- Quant
- Q5_K_M
- Speed (est.)
- ~110 tok/s
The most tokens per second per dollar among cards with room to spare.
Fastest
RTX 509032GB · ~$4,300
- Quant
- Q6_K
- Speed (est.)
- ~165 tok/s
The fastest single GPU in our database for this model.
Mac / unified memory
Ryzen AI Max+ 395128GB unified · ~$3,300
- Quant
- Q8_0
- Speed (est.)
- ~26 tok/s
Lots of memory per dollar and silent, but slower than a GPU with the same capacity.
We may earn a commission when you buy through links on this page. It never changes what we recommend. How we make money
Memory
Gemma 4 26B A4B VRAM requirements
Memory needed at each quantization level. Q4_K_M is the best balance of size and quality for most people; move up to Q5 or Q6 if you have the room.
| Quantization | Quality | Weights | Total @ 8K ctx | Total @ 32K ctx |
|---|---|---|---|---|
| Q2_K | Rough | 9.8 GB | 11 GB | 12 GB |
| Q3_K_M | Usable | 13 GB | 14 GB | 15 GB |
| Q4_K_M | RecommendedDefault | 16 GB | 17 GB | 18 GB |
| Q5_K_M | Very good | 19 GB | 20 GB | 21 GB |
| Q6_K | Near-lossless | 21 GB | 23 GB | 24 GB |
| Q8_0 | Practically lossless | 28 GB | 30 GB | 30 GB |
| BF16 | Full precision | 52 GB | 55 GB | 55 GB |
Context
How context length changes the memory you need
Every token in the context window is stored in the KV cache. At Q4_K_M, with a 16-bit KV cache, Gemma 4 26B A4B needs:
| Context | KV cache | Total memory | With 8-bit KV |
|---|---|---|---|
| 4K tokens | 0.4 GB | 17.2 GB | 17.0 GB |
| 8K tokens | 0.5 GB | 17.3 GB | 17.1 GB |
| 16K tokens | 0.6 GB | 17.5 GB | 17.2 GB |
| 32K tokens | 1.0 GB | 17.8 GB | 17.3 GB |
| 64K tokens | 1.6 GB | 18.5 GB | 17.7 GB |
| 128K tokens | 3.0 GB | 19.8 GB | 18.3 GB |
| 256K tokens | 5.7 GB | 22.5 GB | 19.7 GB |
30 layers in a 5:1 pattern of 1,024-token sliding-window and global layers. Only the global layers grow with context; assuming 5 global layers with 2 KV heads of 512 dims: 2 x 5 x 2 x 512 x 2 bytes = 20,480 bytes per token, about 5.4GB at the full 256K. The sliding layers add a fixed buffer of a few hundred MB.
Compatibility
Every GPU that runs Gemma 4 26B A4B
Each GPU at the highest-quality quantization that fits with 8K context, fastest first. Speeds are single-user estimates.
| GPU | Memory | Quant | Speed | Fit | Price |
|---|---|---|---|---|---|
| RTX 5090 | 32GB | Q6_K | ~165 tok/s | Runs well | ~$4,300 |
| RTX PRO 6000 | 96GB | Q8_0 | ~140 tok/s | Runs well | ~$16,000 |
| RTX 4090 | 24GB | Q5_K_M | ~125 tok/s | Runs well | ~$2,500 used |
| RTX PRO 5000 72GB | 72GB | Q8_0 | ~115 tok/s | Runs well | ~$9,100 |
| RTX 3090 | 24GB | Q5_K_M | ~110 tok/s | Runs well | ~$1,050 used |
| RX 7900 XTX | 24GB | Q5_K_M | ~91 tok/s | Runs well | ~$1,050 used |
| RTX 6000 Ada | 48GB | Q8_0 | ~89 tok/s | Runs well | ~$4,750 used |
| V100 32GB | 32GB | Q6_K | ~84 tok/s | Runs well | ~$700 used |
| M5 Ultra 512GB | 512GB unified | Q8_0 | ~73 tok/s | Runs well | — |
| M5 Ultra 256GB | 256GB unified | Q8_0 | ~73 tok/s | Runs well | ~$9,499 |
| RTX A6000 | 48GB | Q8_0 | ~68 tok/s | Runs well | ~$4,400 used |
| RTX 4000 Ada | 20GB | Q4_K_M | ~63 tok/s | Runs well | $1,250 |
| R9700 | 32GB | Q6_K | ~61 tok/s | Runs well | ~$1,800 |
| MI50 32GB | 32GB | Q6_K | ~61 tok/s | Runs well | ~$600 used |
| M3 Ultra 512GB | 512GB unified | Q8_0 | ~54 tok/s | Runs well | $9,499 |
| M5 Max 128GB | 128GB unified | Q8_0 | ~49 tok/s | Runs well | ~$5,100 |
| Arc Pro B70 | 32GB | Q6_K | ~48 tok/s | Runs well | ~$1,300 |
| M4 Max 128GB | 128GB unified | Q8_0 | ~44 tok/s | Runs well | $3,499 |
| Arc Pro B60 | 24GB | Q5_K_M | ~43 tok/s | Runs well | ~$650 |
| Tesla P40 | 24GB | Q5_K_M | ~36 tok/s | Runs well | ~$230 used |
| Ryzen AI Max+ PRO 495 | 192GB unified | Q8_0 | ~27 tok/s | Runs well | ~$6,800 |
| DGX Spark | 128GB unified | Q8_0 | ~27 tok/s | Runs well | ~$6,950 |
| Ryzen AI Max+ 395 | 128GB unified | Q8_0 | ~26 tok/s | Runs well | ~$3,300 |
| RTX 5080 | 16GB | Q3_K_M | ~155 tok/s | Fits, tight | ~$1,600 |
| RTX 5070 Ti | 16GB | Q3_K_M | ~145 tok/s | Fits, tight | ~$1,200 |
| RTX 4080 Super | 16GB | Q3_K_M | ~130 tok/s | Fits, tight | ~$1,150 used |
| RTX 4070 Ti Super | 16GB | Q3_K_M | ~120 tok/s | Fits, tight | ~$900 used |
| RTX 5060 Ti 16GB | 16GB | Q3_K_M | ~90 tok/s | Fits, tight | ~$620 |
| RX 9070 XT | 16GB | Q3_K_M | ~90 tok/s | Fits, tight | ~$800 |
| RTX 4060 Ti 16GB | 16GB | Q3_K_M | ~63 tok/s | Fits, tight | ~$380 used |
| RX 9060 XT 16GB | 16GB | Q3_K_M | ~54 tok/s | Fits, tight | ~$470 |
With CPU offload (partial offload, assuming 64GB of system RAM): RTX 5070 (~39 tok/s), RTX 4070 (~37 tok/s), RTX 3060 12GB (~33 tok/s), Arc B580 (~30 tok/s), RTX 5060 (~26 tok/s).
Tips
Getting the most out of Gemma 4 26B A4B
- Start at Q4_K_M. It keeps nearly all of the quality at about a third of the 16-bit size. Step up to Q5_K_M or Q6_K only if it still fits with the context you need.
- Budget memory for context. The KV cache grows with every token. Quantizing it to 8-bit roughly halves that cost with little quality loss.
- Offload experts, not layers. For mixture-of-experts models, llama.cpp's
--n-cpu-moe(or-ottensor overrides) keeps attention on the GPU and moves expert weights to system RAM. That keeps speed usable when the model does not fit in VRAM. - Pick the right runtime. Ollama and LM Studio are easiest. llama.cpp gives the most control. vLLM or SGLang are best when several people or agents share one machine.
Want to try different quantizations, context lengths, or GPU counts? Use our VRAM calculator, or read the Gemma 4 26B A4B overview for setup commands and benchmarks.
Frequently asked questions
How much VRAM do I need to run Gemma 4 26B A4B?
About 17GB at Q4_K_M with an 8K context window. At Q8_0 (near-lossless) it needs about 30GB, and at Q3_K_M about 14GB with some quality loss. Longer contexts add 0.06GB per 1K tokens with a 16-bit KV cache.
Can I run Gemma 4 26B A4B on a 16GB GPU?
Yes. A 16GB card like the RTX 5060 Ti 16GB runs it at Q3_K_M at ~155 tokens/sec.
Can I run Gemma 4 26B A4B on a Mac?
Yes. With 128GB of unified memory or more (for example the Ryzen AI Max+ 395), macOS can give the GPU enough memory to run it at Q8_0, at about ~26 tokens/sec. MLX builds are usually fastest on Apple silicon.
Can I run Gemma 4 26B A4B on CPU only, without a GPU?
Yes, if you have at least 32GB of system RAM. On a desktop with dual-channel DDR5, expect roughly 18 tokens/sec, which is better than you might expect because only the active experts are read for each token. A GPU is recommended for interactive use.