Configs › gemma-4-26B-A4B-it tok/s: Strix Halo 78, DGX Spark 62.9
gemma-4-26B-A4B-it tok/s: Strix Halo 78, DGX Spark 62.9
Measured gemma-4-26B-A4B-it speeds: 78 tok/s on Strix Halo (UD-Q4_K_XL, llama.cpp). 11 configs, each linked to its source, snapshot 2026-10-16.
Snapshot 2026-10-16 · 394 configs on AI boxes and 423 on GPU cards · 123 models. Every number is generated from the tracker snapshot and linked to where it was measured; nothing here is typed by hand.
Quick answers
- How fast is gemma-4-26B-A4B-it on Strix Halo?
- 78 tok/s single-stream decode (UD-Q4_K_XL, llama.cpp), the fastest of 4 measured configs on AMD Strix Halo (Ryzen AI Max+ 395), measured by sypherin, source last checked 2026-09-16. source
- How fast is gemma-4-26B-A4B-it on DGX Spark?
- 62.9 tok/s single-stream decode (UD-IQ3_XXS, llama.cpp/cuda), the fastest of 7 measured configs on NVIDIA DGX Spark (GB10), measured by paulplee, source last checked 2026-10-16. source
- How fast is gemma-4-26B-A4B-it on RTX 5090?
- 212.6 tok/s single-stream decode (UD-IQ3_S, llama.cpp/cuda), the fastest of 17 measured configs on NVIDIA RTX 5090 (32GB), measured by paulplee, source last checked 2026-10-16. source
- How fast is gemma-4-26B-A4B-it on H100 SXM?
- 171.7 tok/s single-stream decode (UD-Q8_K_XL, llama.cpp/cuda), the only measured config on NVIDIA H100 SXM (80GB), measured by sypherin, source last checked 2026-10-14. source
- How fast is gemma-4-26B-A4B-it on H100 PCIe?
- 138 tok/s single-stream decode (UD-Q8_K_XL, llama.cpp/cuda), the only measured config on NVIDIA H100 PCIe (80GB), measured by sypherin, source last checked 2026-10-14. source
- How fast is gemma-4-26B-A4B-it on A100 40GB?
- 99.9 tok/s single-stream decode (UD-Q8_K_XL, llama.cpp/cuda), the only measured config on NVIDIA A100 40GB, measured by sypherin, source last checked 2026-10-14. source
Measured configs for gemma-4-26B-A4B-it on AI boxes
Also filed as: Gemma-4-26B-A4B, Gemma-4-26B-A4B-IT.
| Model | Hardware | Quant | Backend | Decode | Context | Trust | Source |
|---|---|---|---|---|---|---|---|
| Gemma-4-26B-A4B | Strix Halo | UD-Q4_K_XL | llama.cpp | 78 tok/s | Owner-submitted | source | |
| gemma-4-26B-A4B-it | DGX Spark | UD-IQ3_XXS | llama.cpp/cuda | 62.9 tok/s | 8192 | Structured table | source |
| gemma-4-26B-A4B-it | DGX Spark | UD-IQ3_S | llama.cpp/cuda | 62.7 tok/s | 8192 | Structured table | source |
| gemma-4-26B-A4B-it | DGX Spark | UD-IQ4_XS | llama.cpp/cuda | 55.8 tok/s | 8192 | Structured table | source |
| gemma-4-26B-A4B-it | DGX Spark | UD-IQ4_NL | llama.cpp/cuda | 55.5 tok/s | 8192 | Structured table | source |
| gemma-4-26B-A4B-it | DGX Spark | MXFP4 | llama.cpp/cuda | 54.9 tok/s | 8192 | Structured table | source |
| gemma-4-26B-A4B-it | Strix Halo | Q4_K_M | llama.cpp/vulkan | 54.2 tok/s (prefill 1323.4) | 512 | Structured table | source |
| Gemma-4-26B-A4B | Strix Halo | Q4_K_M | llama.cpp/vulkan | 48.5 tok/s (prefill 1142) | 512 | Structured table | source |
| gemma-4-26B-A4B-it | DGX Spark | Q8_0 | llama.cpp/cuda | 47.5 tok/s | 8192 | Structured table | source |
| gemma-4-26B-A4B-it | Strix Halo | UD-Q8_K_XL | llama.cpp/rocm-7.2.4 | 41.8 tok/s | Structured table | source | |
| Gemma-4-26B-A4B-IT | DGX Spark | BF16 | vllm | 22.3 tok/s | Structured table | source |
gemma-4-26B-A4B-it on GPU cards, ranked on their own
GPU cards run the model from their own VRAM: very fast on what fits, but most cannot load what a 128 GB AI box can. Ranked on their own, never against the boxes.
| Model | Hardware | Quant | Backend | Decode | Context | Trust | Source |
|---|---|---|---|---|---|---|---|
| gemma-4-26B-A4B-it | RTX 5090 | UD-IQ3_S | llama.cpp/cuda | 212.6 tok/s | 8192 | Structured table | source |
| gemma-4-26B-A4B-it | RTX 5090 | UD-IQ3_XXS | llama.cpp/cuda | 212.3 tok/s | 8192 | Structured table | source |
| gemma-4-26B-A4B-it | RTX 5090 | UD-Q3_K_M | llama.cpp/cuda | 205.8 tok/s | 8192 | Structured table | source |
| gemma-4-26B-A4B-it | RTX 5090 | UD-Q3_K_XL | llama.cpp/cuda | 201.6 tok/s | 8192 | Structured table | source |
| gemma-4-26B-A4B-it | RTX 5090 | UD-IQ4_XS | llama.cpp/cuda | 199 tok/s | 8192 | Structured table | source |
| gemma-4-26B-A4B-it | RTX 5090 | UD-IQ4_NL | llama.cpp/cuda | 197.7 tok/s | 8192 | Structured table | source |
| gemma-4-26B-A4B-it | RTX 5090 | MXFP4 | llama.cpp/cuda | 196.3 tok/s | 8192 | Structured table | source |
| gemma-4-26B-A4B-it | RTX 5090 | UD-Q4_K_S | llama.cpp/cuda | 195.9 tok/s | 8192 | Structured table | source |
| gemma-4-26B-A4B-it | RTX 5090 | UD-Q4_K_XL | llama.cpp/cuda | 195.5 tok/s | 8192 | Structured table | source |
| gemma-4-26B-A4B-it | RTX 5090 | UD-Q4_K_M | llama.cpp/cuda | 195.1 tok/s | 8192 | Structured table | source |
| gemma-4-26B-A4B-it | RTX 5090 | UD-Q5_K_S | llama.cpp/cuda | 194.3 tok/s | 8192 | Structured table | source |
| gemma-4-26B-A4B-it | RTX 5090 | UD-Q5_K_XL | llama.cpp/cuda | 190.5 tok/s | 8192 | Structured table | source |
| gemma-4-26B-A4B-it | RTX 5090 | UD-Q5_K_M | llama.cpp/cuda | 189.2 tok/s | 8192 | Structured table | source |
| gemma-4-26B-A4B-it | RTX 5090 | UD-Q6_K_XL | llama.cpp/cuda | 188.4 tok/s | 8192 | Structured table | source |
| gemma-4-26B-A4B-it | RTX 5090 | UD-Q6_K | llama.cpp/cuda | 188.1 tok/s | 8192 | Structured table | source |
| gemma-4-26B-A4B-it | RTX 5090 | UD-Q8_K_XL | llama.cpp/cuda | 183.7 tok/s | 8192 | Structured table | source |
| gemma-4-26B-A4B-it | RTX 5090 | Q8_0 | llama.cpp/cuda | 182.1 tok/s | 8192 | Structured table | source |
| gemma-4-26B-A4B-it | H100 SXM | UD-Q8_K_XL | llama.cpp/cuda | 171.7 tok/s | 8192 | Verified (ran it) | source |
| gemma-4-26B-A4B-it | H100 PCIe | UD-Q8_K_XL | llama.cpp/cuda | 138 tok/s | 8192 | Verified (ran it) | source |
| gemma-4-26B-A4B-it | A100 40GB | UD-Q8_K_XL | llama.cpp/cuda | 99.9 tok/s | 8192 | Verified (ran it) | source |
More: all configs · hardware · news · contributors · methodology · llms.txt · llms-full.txt · API (OpenAPI) · JSON snapshot. Built by Altronis, private on-prem AI, Singapore.