Configs › gemma-4-26B-A4B-it › gemma-4-26B-A4B-it on DGX Spark: 69.9 tok/s, 8 measured configs
gemma-4-26B-A4B-it on DGX Spark: 69.9 tok/s, 8 measured configs
Measured gemma-4-26B-A4B-it speeds on DGX Spark: fastest 69.9 tok/s single-stream (Q4_K_M, llama.cpp/cuda), 8 configs with quant, backend and context, 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-4-26B-A4B-it on DGX Spark: 69.9 tok/s (Q4_K_M, llama.cpp/cuda), snapshot 2026-10-29.
Every measured single-stream run of gemma-4-26B-A4B-it on DGX Spark
| Decode tok/s | Quant | Backend | Context | Trust | Measured by | Source checked | Source |
|---|---|---|---|---|---|---|---|
| 69.9 | Q4_K_M | llama.cpp/cuda | Structured table | shamily | 2026-10-26 | source | |
| 62.9 | UD-IQ3_XXS | llama.cpp/cuda | 8192 | Structured table | paulplee | 2026-10-29 | source |
| 62.7 | UD-IQ3_S | llama.cpp/cuda | 8192 | Structured table | paulplee | 2026-10-29 | source |
| 55.8 | UD-IQ4_XS | llama.cpp/cuda | 8192 | Structured table | paulplee | 2026-10-29 | source |
| 55.5 | UD-IQ4_NL | llama.cpp/cuda | 8192 | Structured table | paulplee | 2026-10-29 | source |
| 54.9 | MXFP4 | llama.cpp/cuda | 8192 | Structured table | paulplee | 2026-10-29 | source |
| 47.5 | Q8_0 | llama.cpp/cuda | 8192 | Structured table | paulplee | 2026-10-29 | source |
| 22.3 | BF16 | vllm | Structured table | jvr0x | 2026-10-29 | source |
Best quant on DGX Spark
Fastest plain-decoding run per quant, versus the fastest quant.
- Q4_K_M: 69.9 tok/s, fastest (llama.cpp/cuda)
- UD-IQ3_XXS: 62.9 tok/s, 10% lower (llama.cpp/cuda)
- UD-IQ3_S: 62.7 tok/s, 10% lower (llama.cpp/cuda)
- UD-IQ4_XS: 55.8 tok/s, 20% lower (llama.cpp/cuda)
- UD-IQ4_NL: 55.5 tok/s, 21% lower (llama.cpp/cuda)
- MXFP4: 54.9 tok/s, 21% lower (llama.cpp/cuda)
- Q8_0: 47.5 tok/s, 32% lower (llama.cpp/cuda)
- BF16: 22.3 tok/s, 68% lower (vllm)
Backends on DGX Spark
Fastest plain-decoding run per backend. Quants can differ between rows.
- llama.cpp/cuda: 69.9 tok/s (Q4_K_M)
- vllm: 22.3 tok/s (BF16)
gemma-4-26B-A4B-it on other boxes
- Strix Halo: 78 tok/s best of 5 configs
Questions and answers
- How fast is gemma-4-26B-A4B-it on DGX Spark?
- The fastest of 8 measured configs is 69.9 tok/s single-stream (Q4_K_M, llama.cpp/cuda). The slowest is 22.3 tok/s.
- Which quant of gemma-4-26B-A4B-it is fastest on DGX Spark?
- Q4_K_M at 69.9 tok/s, ahead of UD-IQ3_XXS at 62.9, UD-IQ3_S at 62.7, UD-IQ4_XS at 55.8, UD-IQ4_NL at 55.5, MXFP4 at 54.9, Q8_0 at 47.5, BF16 at 22.3. Plain decoding only.
- Which backend is fastest for gemma-4-26B-A4B-it on DGX Spark?
- llama.cpp/cuda at 69.9 tok/s, then vllm at 22.3. Each backend figure is its best plain-decoding row and may use a different quant.
- How does DGX Spark compare with other boxes for gemma-4-26B-A4B-it?
- DGX Spark best 69.9 tok/s. Other boxes, best measured: Strix Halo 78. Backends and quants differ, so this is the best known number per box, not a controlled test.
See also: gemma-4-26B-A4B-it on every box · DGX Spark hardware page (from S$7,452 in Singapore today).
Last updated: snapshot 2026-10-29; newest source check 2026-10-29.
More: all configs · hardware · news · contributors · methodology · llms.txt · llms-full.txt · API (OpenAPI) · JSON snapshot. Built by Altronis, private on-prem AI, Singapore.