Configs › Gemma-4-31B-IT tok/s: Mac Ultra 13.6, DGX Spark 11
Gemma-4-31B-IT tok/s: Mac Ultra 13.6, DGX Spark 11
Measured Gemma-4-31B-IT speeds: 13.6 tok/s on Mac Ultra (BF16, mlx); 11 tok/s on DGX Spark. 4 configs, each linked to its source, snapshot 2026-10-29.
Snapshot 2026-10-29 · 413 configs on AI boxes and 288 on GPU cards · 121 models. Every number is generated from the tracker snapshot and linked to where it was measured; nothing here is typed by hand.
Fastest measured single-stream run of Gemma-4-31B-IT: 13.6 tok/s on Mac Ultra (BF16, mlx), snapshot 2026-10-29.
Fastest measured run on each box
| Hardware | Decode tok/s | Quant | Backend | Context | Trust | Configs on this box | Source |
|---|---|---|---|---|---|---|---|
| Mac Ultra | 13.6 | BF16 | mlx | 32000 | Extracted, source-linked | 1 | source |
| DGX Spark | 11 | Q4_K_M | llama.cpp/cuda | Structured table | 3 | source |
Best quant per box
Where a box has two or more quants measured for Gemma-4-31B-IT, each quant's fastest plain-decoding run is listed against the fastest. Speculative-decoding runs are left out because their speed depends on the prompt.
DGX Spark
| Quant | Decode tok/s | Versus fastest | Backend | Context | Trust | Source |
|---|---|---|---|---|---|---|
| Q4_K_M | 11 | fastest | llama.cpp/cuda | Structured table | source | |
| NVFP4 | 6.6 | 40% lower | vllm | Structured table | source |
Gemma-4-31B-IT on one box
- Gemma-4-31B-IT on DGX Spark: 3 measured configs, fastest 11 tok/s
Questions and answers
- How fast is Gemma-4-31B-IT on local hardware?
- The fastest measured single-stream run is 13.6 tok/s on Mac Ultra (BF16, mlx). TokenMark tracks 4 single-stream configs for Gemma-4-31B-IT on 2 boxes.
- Which hardware runs Gemma-4-31B-IT fastest?
- Best measured single-stream tok/s per box: Mac Ultra 13.6, DGX Spark 11. Backends and quants differ between boxes, so read this as the best known number per box, not a controlled test.
- Which quant of Gemma-4-31B-IT is fastest on DGX Spark?
- Q4_K_M at 11 tok/s, the fastest of 2 quants measured on DGX Spark. The slowest, NVFP4, reaches 6.6 tok/s (40% lower). Plain decoding only; speculative-decoding runs are left out.
- How many measured configs does TokenMark have for Gemma-4-31B-IT?
- 4 single-stream configs on 2 boxes, newest source check 2026-10-29. Every row links to where it was measured.
All measured configs for Gemma-4-31B-IT
Also filed as: Gemma-4-31B.
| Model | Hardware | Quant | Backend | Decode | Context | Trust | Source |
|---|---|---|---|---|---|---|---|
| Gemma-4-31B | Mac Ultra | BF16 | mlx | 13.6 tok/s (prefill 1217) | 32000 | Extracted, source-linked | source |
| Gemma-4-31B-IT | DGX Spark | Q4_K_M | llama.cpp/cuda | 11 tok/s (prefill 743) | Structured table | source | |
| Gemma-4-31B-IT | DGX Spark | Q4_K_M | llama.cpp/cuda | 10.7 tok/s (prefill 746.3) | Structured table | source | |
| Gemma-4-31B-IT | DGX Spark | NVFP4 | vllm | 6.6 tok/s | Structured table | source |
Last updated: snapshot 2026-10-29; newest source check 2026-10-29. Rankings use single-stream runs only; aggregate and multi-GPU rows are kept off them.
More: all configs · hardware · news · contributors · methodology · llms.txt · llms-full.txt · API (OpenAPI) · JSON snapshot. Built by Altronis, private on-prem AI, Singapore.