Radeon RX 7900 XT
AMD
CompatibleRuns Gemma 4 26B A4B IT Q4_K_M entirely in memory
Model · Google
Gemma 4 26B A4B IT: VRAM and hardware requirements
At Q4_K_M with an 8K context, Gemma 4 26B A4B IT needs about 16.9 GB: a GPU with 20 GB or more runs it entirely in VRAM.
- Parameters25.8B
- Active3.8B
- Max context256K
- LicenseApache-2.0
Specifications
- Organization
- Architecture
- gemma4
- Released
- March 2026
- Layers
- 30
- Hidden size
- 2816
- Attention heads
- 16
- KV heads
- 8
- Head dimension
- 256
- Vocabulary
- 262,144
- KV cache per 1K tokens (f16)
- 0.215 GB
- Repository
- google/gemma-4-26B-A4B-it
Official specData updated todaySource: Hugging Face config.json + model card
Quantizations and memory
| Quantization | File size | Memory needed | Minimum GPU memory | Quality |
|---|---|---|---|---|
| BF16 | 47 GB | ~51.9 GB | 80 GB | Reference (unquantized) |
| Q8_0 | 25 GB | ~28.2 GB | 32 GB | Near-lossless |
| Q6_K | 21.3 GB | ~22.1 GB | 32 GB | Near-lossless |
| Q5_K_M | 18 GB | ~19.4 GB | 24 GB | Very good |
| Q4_K_MMost used | 15.9 GB | ~16.9 GB | 20 GB | Good — the usual choice |
| IQ4_XS | 13.2 GB | ~15.5 GB | 16 GB | Good — the usual choice |
| Q3_K_M | 12.1 GB | ~14 GB | 16 GB | Noticeable loss |
| Q2_K | 10.2 GB | ~11.4 GB | 16 GB | Significant loss |
File sizes: not available until imported from the model repository. Memory = weights (estimated from bits per weight) + KV cache + runtime overhead + safety margin. (BF16 ≈ 16 bpw, Q8_0 ≈ 8.5 bpw, Q6_K ≈ 6.57 bpw, Q5_K_M ≈ 5.7 bpw, Q4_K_M ≈ 4.9 bpw, IQ4_XS ≈ 4.46 bpw, Q3_K_M ≈ 4 bpw, Q2_K ≈ 3.17 bpw)
Q4, Q5 or Q8: how to choose a quantization
GPUs that run Gemma 4 26B A4B IT
| RTX 509032 GB · 64 GB RAM~18.1 GB · 199–334 tok/s estimate | Excellent |
|---|---|
| RTX 409024 GB · 64 GB RAM~18.1 GB · 146–228 tok/s estimate | Excellent |
| RTX 309024 GB · 64 GB RAM~18.1 GB · 139–215 tok/s estimate | Excellent |
| RX 7900 XTX24 GB · 64 GB RAM~18.1 GB · 100–171 tok/s estimate | Excellent |
| RTX 5060 Ti 16 GB16 GB · 64 GB RAM~18.1 GB · 31–58 tok/s estimate | CPU/GPU offload |
| RTX 5070 Ti16 GB · 64 GB RAM~18.1 GB · 36–71 tok/s estimate | CPU/GPU offload |
| RTX 508016 GB · 64 GB RAM~18.1 GB · 37–72 tok/s estimate | CPU/GPU offload |
| RTX 4060 Ti 16 GB16 GB · 64 GB RAM~18.1 GB · 26–48 tok/s estimate | CPU/GPU offload |
| RTX 4070 Ti SUPER16 GB · 64 GB RAM~18.1 GB · 34–66 tok/s estimate | CPU/GPU offload |
| RX 9070 XT16 GB · 64 GB RAM~18.1 GB · 29–59 tok/s estimate | CPU/GPU offload |
| RTX 3060 12 GB12 GB · 64 GB RAM~18.1 GB · 17–31 tok/s estimate | CPU/GPU offload |
| RTX 407012 GB · 64 GB RAM~18.1 GB · 17–33 tok/s estimate | CPU/GPU offload |
| RTX 507012 GB · 64 GB RAM~18.1 GB · 18–34 tok/s estimate | CPU/GPU offload |
| RTX 40608 GB · 64 GB RAM~18.1 GB · 11–21 tok/s estimate | Possible but slow |
Macs that run Gemma 4 26B A4B IT
| M3 Ultra (80-core GPU) · 512 GBMac Studio~17.9 GB · 78–135 tok/s estimate | Excellent |
|---|---|
| M4 Max (40-core GPU) · 128 GBMacBook Pro 14, MacBook Pro 16~17.9 GB · 64–106 tok/s estimate | Excellent |
| M3 Max (40-core GPU) · 128 GBMacBook Pro 14, MacBook Pro 16~17.9 GB · 54–86 tok/s estimate | Excellent |
| M4 Pro (20-core GPU) · 64 GBMac mini, MacBook Pro 14~17.9 GB · 42–64 tok/s estimate | Excellent |
| M4 Max (32-core GPU) · 36 GBMacBook Pro 14, MacBook Pro 16~17.9 GB · 55–87 tok/s estimate | Excellent |
| M4 · 32 GBMacBook Air, MacBook Pro 14~17.9 GB · 22–32 tok/s estimate | Excellent |
| M5 · 32 GBMacBook Pro 14~17.9 GB · 27–40 tok/s estimate | Excellent |
| M3 · 24 GBMacBook Air, MacBook Pro 14~17.9 GB · No data | Incompatible |
EstimateData updated 40 days agoSource: Run LLM Locally engine
Limits in plain words
Gemma 4 26B A4B IT is a mixture-of-experts model: all 25.8B parameters must be in memory, but only 3.8B are read per token, so it generates much faster than a dense model of the same size.
The KV cache grows with context: at 32K it adds about 0.8 GB on top of the weights.
With under 8 GB of VRAM, use a smaller quantization or a smaller model of the same family.
Where to buy
Prices and links follow the market you choose.
AMD Ryzen AI Max+ 395
Unified memory PC
ExcellentRuns Gemma 4 26B A4B IT Q4_K_M entirely in memory
No current offer imported for this market: the link opens a marketplace search.
Frequently asked questions
How much VRAM does Gemma 4 26B A4B IT need?
About 16.9 GB at Q4_K_M with an 8K context, including the KV cache and runtime overhead. At Q8_0 it needs about 28.2 GB. A longer context adds more: about 0.8 GB of KV cache at 32K.
Can Gemma 4 26B A4B IT run on a Mac?
Yes on a Mac whose GPU can use at least 17.9 GB, e.g. M4 32 GB. macOS lets the GPU use about two thirds to three quarters of unified memory by default.
How fast is Gemma 4 26B A4B IT?
On a RTX 4090, we estimate 146–228 tok/s at Q4_K_M. This is an estimate from memory bandwidth (confidence low) until benchmarks are imported.