Run Gemma 4 26B-A4B locally: memory, speed and cost
Gemma 4 26B-A4B has 25.2 billion parameters, 3.8 billion active per token (mixture of experts). At Q4_K_M it needs about 18 GB with 32K tokens of context, so the cheapest hardware here that runs it is the Intel Arc Pro B70 (about 77 tokens a second). It scores 142 on Epoch AI’s Capabilities Index. Through an API it costs $0.042 / $0.22 per 1M tokens.
Memory needed
Weights plus context cache
| Quantisation | 8K context | 32K context | 128K context |
|---|---|---|---|
| Q4_K_Musual choice | 17 GB | 18 GB | 22 GB |
| Q5_K_M | 20 GB | 21 GB | 25 GB |
| Q8_0 | 29 GB | 30 GB | 34 GB |
| BF16 | 53 GB | 54 GB | 58 GB |
Estimates: weights at the stated bits per weight, plus the context (KV) cache, plus about 1 GB of overhead. Lower-bit quantisations save memory at some cost in quality.
Gemma 4 26B-A4B on each device
Q4_K_M, 32K context
| Hardware | Memory | Fits? | Speed (estimate) |
|---|---|---|---|
| RTX 5090 | 32 GB | Fits | 204 tok/s |
| RTX 5080 | 16 GB | Partly, with system RAM | 63 tok/s |
| RTX 5070 Ti | 16 GB | Partly, with system RAM | 61 tok/s |
| RTX 5060 Ti 16 GB | 16 GB | Partly, with system RAM | 48 tok/s |
| RTX 4090 (used) | 24 GB | Fits | 145 tok/s |
| RTX 3090 (used) | 24 GB | Fits | 138 tok/s |
| RTX 3060 12 GB | 12 GB | Partly, with system RAM | 31 tok/s |
| RTX PRO 6000 Blackwell | 96 GB | Fits | 204 tok/s |
| Radeon RX 9070 XT | 16 GB | Partly, with system RAM | 55 tok/s |
| Radeon RX 7900 XTX | 24 GB | Fits | 140 tok/s |
| Radeon AI PRO R9700 | 32 GB | Fits | 112 tok/s |
| Intel Arc Pro B70 | 32 GB | Fits | 77 tok/s |
| Mac Studio M5 Max 128 GB | 128 GB | Fits | 99 tok/s |
| Mac Studio M5 Ultra 512 GB | 512 GB | Fits | 118 tok/s |
| Ryzen AI Max+ 395 128 GB | 128 GB | Fits | 46 tok/s |
| NVIDIA DGX Spark | 128 GB | Fits | 49 tok/s |
Generation speed estimated from memory bandwidth, calibrated against published llama.cpp results. Fastest here: RTX 5090, about 204 tokens a second. Try other settings in the calculator.
How capable is it?
Capability 142 (range 139–144) on the Epoch Capabilities Index. API models with a similar score: Gemini 3.5 Flash-Lite (145, $0.30 / $2.50) and MiniMax M3 (147, $0.30 / $1.20). Similar local models: Gemma 4 31B (143), Qwen3.6 35B-A3B (144), gpt-oss-120b (140).
Source: Epoch Capabilities Index by Epoch AI, CC BY 4.0, read 4 Nov 2026.
Or use it through an API
$0.042 / $0.22 per 1M input / output tokens through OpenRouter (Gemma 4 26B-A4B hosted prices). A customer chat assistant of 100M input and 20M output tokens a month costs about $8.60, so local hardware pays for itself only with heavy use.
Questions
How much VRAM does Gemma 4 26B-A4B need?
About 17 GB at Q4_K_M with 8K tokens of context, 18 GB with 32K and 22 GB with 128K. At 8-bit (Q8_0) it needs about 29 GB with 8K context.
Can Gemma 4 26B-A4B run on a Mac?
Yes. A Mac Studio M5 Max with 128GB runs it with 32K context at an estimated 99 tokens a second.
How fast is Gemma 4 26B-A4B on an RTX 4090?
About 145 tokens a second with 32K context, an estimate from the card’s memory bandwidth.
How capable is Gemma 4 26B-A4B?
It scores 142 on Epoch AI’s Capabilities Index (range 139–144), close to Gemini 3.5 Flash-Lite (145) and MiniMax M3 (147) among API models.
What does Gemma 4 26B-A4B cost through an API?
$0.042 / $0.22 per 1M input / output tokens through OpenRouter, the cheapest host we track. A customer chat assistant of 100M input and 20M output tokens a month would cost about $8.60.