DiffusionGemma 26B A4B needs ~21.1 GB VRAM. Intel Arc Pro B60 24GB has 24.0 GB. With Q3_K_S quantization, expect ~34 tok/s.
Operating mode
Interactive favors responsiveness, while light API and scale-out lean harder on serving readiness. The fit stays the same, but the recommendation lens changes.
Current mode
Balanced
Balanced for general local use. Keeps the ranking neutral across personal and serving workflows.
Select quantization to explore
0.2 GB over capacity — needs offload or smaller quantization
Fit status
Too heavy
Decode
22.1 tok/s
TTFT
8745 ms
Safe context
15K
Memory
24.2 GB / 24.0 GB
The raw memory story may look fine, but the software ecosystem is still a constraint here.
Runtime ecosystem is narrower than CUDA
Intel GPUs can look attractive on memory per dollar, but local AI tooling, kernels, and model coverage are still broader and easier on CUDA today.
Prefer CUDA if you want the path of least resistance
If your goal is maximum runtime coverage, easier troubleshooting, and better support for new local AI releases, CUDA is usually still the safer upgrade path.
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Chat | A | Tight fit | 29.4 tok/s | 3586 ms | 15K |
| Coding | F | Too heavy | 22.1 tok/s | 8745 ms | 15K |
| Agentic Coding | F | Too heavy | 16.6 tok/s | 16956 ms | 15K |
| Reasoning | F | Too heavy | 22.1 tok/s | 10335 ms | 15K |
| RAG | F | Too heavy | 16.6 tok/s | 21195 ms | 15K |
Inference speed
Estimated decode speed (tokens/sec) for DiffusionGemma 26B A4B at Q4_K_M across popular GPUs and Apple Silicon, using the fastest local runtime per device. Fastest is RTX 5090 32GB at ~144 tok/s. Speed is memory-bandwidth bound, so cards that fit the whole model in VRAM run far faster than ones that offload to system RAM.
| GPU / Mac | Memory | Quant | Speed (tok/s) | Fits? |
|---|---|---|---|---|
 RTX 5090 32GB | 32 GB | Q4_K_M | 143.6 | Fits |
| RTX 4090 24GB | 24 GB | Q4_K_M | 67.5 | Too big |
Mac Studio M3 Ultra 256GB | 256 GB | Q4_K_M | 66.6 | Fits |
RX 7900 XTX 24GB | 24 GB | Q4_K_M | 60.9 | Too big |
| RTX 3090 24GB | 24 GB | Q4_K_M | 57.7 | Too big |
Mac Studio M2 Ultra 128GB | 128 GB | Q4_K_M | 55.5 | Fits |
Mac Studio M1 Ultra 128GB | 128 GB | Q4_K_M | 52.6 | Fits |
MacBook Pro M4 Max 128GB | 128 GB | Q4_K_M | 41.1 | Fits |
MacBook Pro M4 Max 64GB | 64 GB | Q4_K_M | 41.1 | Fits |
MacBook Pro M3 Max 64GB | 64 GB | Q4_K_M | 28.7 | Fits |
MacBook Pro M1 Max 64GB | 64 GB | Q4_K_M | 26.3 | Fits |
MacBook Pro M4 Pro 48GB | 48 GB | Q4_K_M | 25.1 | Fits |
| RTX 4080 Super 16GB | 16 GB | Q4_K_M | 24.6 | Too big |
| RTX 4070 12GB | 12 GB | Q4_K_M | 8.6 | Too big |
| RTX 3060 12GB | 12 GB | Q4_K_M | 5.4 | Too big |
| RTX 4060 8GB | 8 GB | Q4_K_M | 3.6 | Too big |
Estimates for single-stream decoding at Q4_K_M; real tokens/sec varies with prompt length, context, batch size, and runtime build. Prompt processing (prefill) is faster than the decode figures shown here.
How DiffusionGemma 26B A4B (25.799999237060547B params) fits at each quantization level on Intel Arc Pro B60 24GB (24.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 10.1 GB | Low | A78 |
Q3_K_S | 3 | 12.6 GB | Low | A79 |
NVFP4 | 4 | 14.4 GB | Medium | A79 |
Q4_K_M | 4 | 15.7 GB | Medium | A78 |
Q5_K_MBest for your GPU | 5 | 18.6 GB | High | A78 |
Q6_K | 6 | 21.2 GB | High | F0 |
Q8_0 | 8 | 27.6 GB | Very High | F0 |
F16 | 16 | 52.9 GB | Maximum | F0 |
Copy-paste commands to run DiffusionGemma 26B A4B on your machine.
Run
docker run --rm -it ghcr.io/ggerganov/llama.cpp:full \
--hf-repo "google/diffusiongemma-26B-A4B-it" \
--hf-file "diffusiongemma-26B-A4B-it-Q4_K_M.gguf" \
-c 4096 -ngl 99Upgrade-Optionen
RTX 5090 32GBGrößter SprungMakes the model fit on the accelerator instead of staying completely out of reach.
Raises estimated decode speed by about 550%.
ca. $1,999 MSRP
Intel Data Center GPU Max 1550 128GBGünstige WahlMakes the model fit on the accelerator instead of staying completely out of reach.
Raises estimated decode speed by about 991%.
ca. $15,000 MSRP
Gaudi 3 128GBBestes Preis-Leistungs-VerhältnisMakes the model fit on the accelerator instead of staying completely out of reach.
Raises estimated decode speed by about 1301%.
ca. $15,000 MSRP
Yes, Intel Arc Pro B60 24GB can run DiffusionGemma 26B A4B at Q3_K_S quantization (Tight fit). The recommended Q4_K_M requires 24.2 GB which exceeds available memory, but at Q3_K_S it needs only 21.1 GB. Expected decode speed: 34.1 tok/s.
DiffusionGemma 26B A4B (25.799999237060547B parameters) requires approximately 24.2 GB at Q4_K_M quantization. On Intel Arc Pro B60 24GB, it fits at Q3_K_S using 21.1 GB.
The recommended quantization is Q4_K_M, but on Intel Arc Pro B60 24GB the best fitting quantization is Q3_K_S, which uses 21.1 GB.
On Intel Arc Pro B60 24GB, DiffusionGemma 26B A4B achieves approximately 34.1 tokens per second decode speed with a time-to-first-token of 5679ms using Q3_K_S quantization.
For coding workloads, DiffusionGemma 26B A4B on Intel Arc Pro B60 24GB receives a F grade with 22.1 tok/s and 15K context.
On Intel Arc Pro B60 24GB, DiffusionGemma 26B A4B can safely use up to 29K tokens of context at Q3_K_S quantization. The model's official context limit is 262K, but available memory constrains the safe maximum.
Prefer CUDA if you want the path of least resistance. If your goal is maximum runtime coverage, easier troubleshooting, and better support for new local AI releases, CUDA is usually still the safer upgrade path.
Often yes, if your goal is the easiest setup and the widest runtime support. Intel can offer attractive memory capacity, but CUDA still tends to win on tooling maturity, guides, kernels, and model coverage for local AI.
Paste this snippet into any page to show a live fit card.
<iframe src="https://willitrunai.com/embed/diffusiongemma-26b-a4b-on-arc-pro-b60-24gb" width="400" height="180" frameborder="0" style="border:none;border-radius:12px;overflow:hidden;" title="Will It Run AI — fit result"></iframe>
Preview: