Gemma 4 26B-A4B (MoE)
Gemma 4 26B-A4B (MoE) needs 16.4 GB of VRAM at Q4_K_M with 8k context, computed from its config.json as checked on 3 November 2026. The smallest card that holds it has 24 GB, and the fastest desktop card that runs it, the RTX 5090, generates at most 569 tokens per second.
- Parameters
- 25.8B, 3.8B active
- Context
- 262,144 tokens
- Licence
- Apache 2.0
- Released
- March 2026
- Config checked
- 3 November 2026
Open Gemma 4 26B-A4B (MoE) in the LLM checker
Which card size runs it
16.4 GB against each memory size. A model may use 95% of a card, and only what is left over, or missing, is coloured; tight means above 85%, with little room for a longer conversation.
Other quantisations, context lengths or a Q8 KV cache: the VRAM calculator.
Computed 3 November 2026Which GPU to pick
The smallest card here that loads it at Q4_K_M with 8k of context is the RX 7900 XT (20 GB): 16.4 GB needed, an estimated 254 tokens per second. The same card still has room at 32,768 tokens of context (16.9 GB), so there is no reason to buy above it for this model.
At Q8_0, which is near-lossless, the card is the RTX 5090 (32 GB), needing 29.3 GB. On Apple silicon, a Mac mini M4 Pro holds it with 32,768 tokens of context at an estimated 87 tokens per second; prompt processing is slower than on Nvidia, which long documents make obvious. Among unified-memory desktops, the Ryzen AI Max+ 395 holds it at 32,768 tokens in 16.9 GB, at an estimated 81 tokens per second.
Speed on each card
| GPU | Fits | Est. tokens/s | Feels like | Memory | Bandwidth |
|---|---|---|---|---|---|
| H100 SXM | Yes | ≤ 1,064 | faster than you can read | 80 GB | 3,350 GB/s |
| A100 | Yes | ≤ 648 | faster than you can read | 80 GB | 2,039 GB/s |
| RTX 5090 | Yes | ≤ 569 | faster than you can read | 32 GB | 1,792 GB/s |
| RTX PRO 6000 Blackwell | Yes | ≤ 569 | faster than you can read | 96 GB | 1,792 GB/s |
| RTX PRO 5000 Blackwell | Yes | ≤ 427 | faster than you can read | 48 GB | 1,344 GB/s |
| RTX 4090 | Yes | ≤ 320 | faster than you can read | 24 GB | 1,008 GB/s |
| RX 7900 XTX | Yes | ≤ 305 | faster than you can read | 24 GB | 960 GB/s |
| RTX 6000 Ada | Yes | ≤ 305 | faster than you can read | 48 GB | 960 GB/s |
| RTX 3090 | Yes | ≤ 297 | faster than you can read | 24 GB | 936 GB/s |
| RTX 5090 Laptop GPU | Yes | ≤ 285 | faster than you can read | 24 GB | 896 GB/s |
| RTX PRO 4500 Blackwell | Yes | ≤ 285 | faster than you can read | 32 GB | 896 GB/s |
| L40S | Yes | ≤ 274 | faster than you can read | 48 GB | 864 GB/s |
| RX 7900 XT | Yes | ≤ 254 | faster than you can read | 20 GB | 800 GB/s |
| RTX A6000 | Yes | ≤ 244 | faster than you can read | 48 GB | 768 GB/s |
| RTX PRO 4000 Blackwell | Yes | ≤ 213 | faster than you can read | 24 GB | 672 GB/s |
| Radeon AI PRO R9700 | Yes | ≤ 203 | faster than you can read | 32 GB | 640 GB/s |
| Arc Pro B70 | Yes | ≤ 193 | faster than you can read | 32 GB | 608 GB/s |
| Arc Pro B60 | Yes | ≤ 145 | faster than you can read | 24 GB | 456 GB/s |
| RTX PRO 4000 Blackwell SFF | Yes | ≤ 137 | faster than you can read | 24 GB | 432 GB/s |
| RTX Spark | Yes | 95 | faster than you can read | 128 GB (102 for the GPU) | 300 GB/s |
| DGX Spark | Yes | 87 | faster than you can read | 128 GB (126 for the GPU) | 273 GB/s |
| Ryzen AI Max+ 395 | Yes | 81 | faster than you can read | 128 GB (96 for the GPU) | 256 GB/s |
| RTX 4060 Ti | No | - | does not fit | 16 GB | 288 GB/s |
| RTX 5060 Ti | No | - | does not fit | 16 GB | 448 GB/s |
| RTX 5080 | No | - | does not fit | 16 GB | 960 GB/s |
| RTX 5070 Ti | No | - | does not fit | 16 GB | 896 GB/s |
| RTX 4080 SUPER | No | - | does not fit | 16 GB | 736 GB/s |
| RX 9070 XT | No | - | does not fit | 16 GB | 640 GB/s |
| RX 9060 XT | No | - | does not fit | 16 GB | 320 GB/s |
| RTX 3060 | No | - | does not fit | 12 GB | 360 GB/s |
| RTX 4070 | No | - | does not fit | 12 GB | 504 GB/s |
| RTX 5070 | No | - | does not fit | 12 GB | 672 GB/s |
| Arc B580 | No | - | does not fit | 12 GB | 456 GB/s |
Single-stream decode, from memory bandwidth at 70% efficiency. Prompt processing and batching not included. These are ceilings, not measurements: Gemma 4 26B-A4B (MoE) reads only 2.2 GB of weights per token at Q4, so little that the per-token costs the formula leaves out decide much of the real speed. The 19 figures marked ≤ are where a real runtime falls furthest below the number shown. See the speed estimator for other quantisations and Apple hardware, or every card compared if you are choosing hardware rather than a model.
Computed 3 November 2026Memory by quantisation and context
| Quant | 4,096 ctx | 8,192 ctx | 32,768 ctx |
|---|---|---|---|
| Q8_0 | 29.2 GB · fits 32 GB | 29.3 GB · fits 32 GB | 29.8 GB · fits 32 GB |
| Q5_K_M | 19.8 GB · fits 24 GB | 19.9 GB · fits 24 GB | 20.4 GB · fits 24 GB |
| Q4_K_M | 16.4 GB · fits 24 GB | 16.4 GB · fits 24 GB | 16.9 GB · fits 24 GB |
Weights at this quant: 15 GB at Q4. Every extra 1,000 tokens of context adds about 0.02 GB of KV cache at FP16, on top of a fixed 0.21 GB that the sliding-window layers hold once their window is full. Try other settings in the VRAM calculator.
About the model
Gemma 4 26B-A4B (MoE) has 25.8B parameters with 3.8B active per token (mixture of experts). It has 30 layers: 25 use sliding-window attention over the last 1,024 tokens (8 KV heads of dimension 256), and 5 are global layers with 2 KV heads of dimension 512. The mixture-of-experts member of Gemma 4, spring 2026: 128 experts with 8 active plus one shared, 3.8B parameters read per token. Google positions it as the fast alternative to the dense 31B.
Text and images 262,144-token window Apache 2.0 March 2026 3.8B of 25.8B active
Reasoning. Configurable thinking modes, per the model card.
All 128 experts stay in memory; 3.8B are active per token. The parameter count includes the vision encoder; a text-only GGUF is slightly smaller.
Native context window: 262,144 tokens. Licence: Apache 2.0, a permissive licence. Weights published in March 2026.
Good at
- Speed on a 24 GB card: it answers like a 4B model while holding 26B of weights.
- Long context cheaply: 25 of its 30 layers are sliding-window.
- Apache 2.0.
Watch out for
- All 25.8B parameters must fit in memory even though few are used per token. It is fast, not small.
- A mixture-of-experts model this sparse trades depth for speed. For the hardest reasoning, the dense 31B is the one Google built for it.
- No audio input; in Gemma 4 that is the 12B and the E-series.
Similar models
The models that need about the same memory as Gemma 4 26B-A4B (MoE) at Q4_K_M and 8k, which makes them the real alternatives on whatever card you have. The context column is what each extra 1,000 tokens costs, and it is where models of the same size differ most.
| Model | Size | Needs | Per 1k context | Window | Released |
|---|---|---|---|---|---|
| Gemma 4 26B-A4B (MoE) | 25.8B · 3.8B active | 16.4 GB | 0.02 GB | 256k | 2026-03 |
| Devstral Small 2 24B | 24B | 16.3 GB | 0.164 GB | 256k | 2025-12 |
| Mistral Small 3.1 24B | 24B | 16.3 GB | 0.164 GB | 128k | 2025-03 |
| Qwen3.8 27B | 27.8B | 18 GB | 0.066 GB | 256k | 2026-08 |
| Qwen3.6 27B | 27.8B | 18 GB | 0.066 GB | 256k | 2026-04 |
See every model compared, which models fit on 8 to 96 GB, or the cheapest hour of every cloud card.
Questions
- Can I run Gemma 4 26B-A4B (MoE) on a 24 GB card?
- Yes, at Q4_K_M and 8k context it needs about 16.4 GB, which fits a 24 GB card with room.
- How much VRAM does Gemma 4 26B-A4B (MoE) need at Q8?
- About 29.3 GB at 8k context, or 29.8 GB at 32k. Q8 is near-lossless; use it when it fits.
- What is the context length of Gemma 4 26B-A4B (MoE)?
- 262,144 tokens natively. Holding all of it at Q4_K_M takes about 21.6 GB, of which 5.6 GB is context.
- Why does Gemma 4 26B-A4B (MoE) need so little memory for long context?
- Because most of its layers do not keep a cache that grows. It has 30 layers: 25 use sliding-window attention over the last 1,024 tokens (8 KV heads of dimension 256), and 5 are global layers with 2 KV heads of dimension 512. At 32k tokens the context costs about 0.9 GB; if all 30 layers kept a conventional full-length cache, the same conversation would cost about 8.1 GB. That saving depends on the runtime implementing the sliding-window cache. llama.cpp and vLLM do; a build that does not will use far more.
- How fast is Gemma 4 26B-A4B (MoE) on an RTX 4090?
- At most about 320 tokens per second at Q4, single stream, which is faster than you can read. That is a ceiling worked out from the card's memory bandwidth, not a measurement, and at a figure this high real runtimes land well below it, because the costs the formula leaves out take a growing share of each token.
- Architecture values from config.json in google/gemma-4-26B-A4B-it on Hugging Face, and its model card, checked 3 November 2026. Every figure on this page is computed from those values; check them before buying hardware for this model.
- Estimates assume a single conversation on a card that is otherwise free. A desktop on the same GPU takes 0.5 to 2 GB.