metadata
license: mit
tags:
- benchmark
- llama.cpp
- mixture-of-experts
- gguf
pretty_name: PolyStrata speed measurements
PolyStrata speed measurements
Tokens per second of PolyStrata, an open-source engine that runs mixture-of-experts models on a GPU too small for them by keeping the experts the model uses most on the card and running the rest on the CPU. The numbers are the ones in the project's README at commit b0e3b57a, plus the results people reported on their own cards. How each number is measured is in docs/BACKENDS.md; the benchmark is python polystrata.py benchmark --model <name> --prompt-tokens 24000, and anyone can rerun it.
Files:
readme_speeds.csv: the README's tables. "Writes" is the speed while an answer is written in a short chat; "reads" is the speed at which a 10,000-token prompt is read.community_reports.csv: what people measured on their own machines, with the source of each report.
The README's numbers
| Model | Size | Machine | Writes, tokens/s | Reads, tokens/s |
|---|---|---|---|---|
| Qwen3.8-Flash-Next | Q2_0 | NVIDIA: RTX 5070 (12 GB), Ryzen 5 7600, 64 GB RAM | 94 | 2,650 |
| Qwen3.8-Flash-Next | IQ2_XS | NVIDIA: RTX 5070 (12 GB), Ryzen 5 7600, 64 GB RAM | 79 | 2,090 |
| Qwen3.8-Flash-Next | IQ3_XXS | NVIDIA: RTX 5070 (12 GB), Ryzen 5 7600, 64 GB RAM | 62 | 1,750 |
| Qwen3.8-Flash-Next | IQ3_S | NVIDIA: RTX 5070 (12 GB), Ryzen 5 7600, 64 GB RAM | 53 | 1,620 |
| Qwen3.8-Flash-Next | Coder | NVIDIA: RTX 5070 (12 GB), Ryzen 5 7600, 64 GB RAM | 55 | 2,180 |
| Qwen3.8-Flash-Next | Q2_0 | AMD: RX 9070 XT (16 GB), Ryzen 9 3900X, 47 GB RAM | 60 | 1,160 |
| Qwen3.8-Flash-Next | IQ2_XS | AMD: RX 9070 XT (16 GB), Ryzen 9 3900X, 47 GB RAM | 52 | 1,110 |
| Qwen3.8-Flash-Next | Coder | AMD: RX 9070 XT (16 GB), Ryzen 9 3900X, 47 GB RAM | 44 | 1,420 |
| GLM 5.3 Flash | 3.0BIT | NVIDIA: RTX 5090 (32 GB), Xeon Platinum 8470Q (25 vCPU), 120 GB RAM | 59-66 | 820 |
| GLM 5.3 Flash | UD-IQ2_XXS | NVIDIA: RTX 5090 (32 GB), Xeon Platinum 8470Q (25 vCPU), 120 GB RAM | 50-52 | 1,060-1,080 |
| GLM 5.3 Flash | REAP50-Q3_K_M | NVIDIA: RTX 5090 (32 GB), Xeon Platinum 8470Q (25 vCPU), 120 GB RAM | 59-66 | 1,300-1,310 |
| Qwen3.6-35B-A3B | UD-Q3_K_M | NVIDIA: RTX 4070 Laptop (8 GB), Core i7-14700HX, 32 GB RAM | 94-100 | 1,690 |
| Qwen3.6-35B-A3B | UD-Q2_K_XL | NVIDIA: RTX 4070 Laptop (8 GB), Core i7-14700HX, 32 GB RAM | 123-124 | 1,590 |
| Qwen3.6-35B-A3B | UD-Q4_K_M | NVIDIA: RTX 4070 Laptop (8 GB), Core i7-14700HX, 32 GB RAM | 82-86 | 1,560-1,570 |
| Qwen3.6-35B-A3B | UD-Q3_K_M | NVIDIA: RTX 5090 (32 GB), Xeon Platinum 8470Q (25 vCPU), 120 GB RAM | 394-395 | 6,350-6,550 |
| Qwen3.6-35B-A3B | UD-Q2_K_XL | NVIDIA: RTX 5090 (32 GB), Xeon Platinum 8470Q (25 vCPU), 120 GB RAM | 450 | 6,400-6,590 |
| Qwen3.6-35B-A3B | UD-Q4_K_M | NVIDIA: RTX 5090 (32 GB), Xeon Platinum 8470Q (25 vCPU), 120 GB RAM | 363-401 | 6,030-6,690 |
| Ornith 1.5 | APEX-I-COMPACT | NVIDIA: RTX 4070 Laptop (8 GB), Core i7-14700HX, 32 GB RAM | 76-82 | 1,620 |
| Ornith 1.5 | APEX-I-MINI | NVIDIA: RTX 4070 Laptop (8 GB), Core i7-14700HX, 32 GB RAM | 80-86 | 1,550-1,560 |
| Ornith 1.5 | APEX-I-QUALITY | NVIDIA: RTX 4070 Laptop (8 GB), Core i7-14700HX, 32 GB RAM | 59-63 | 1,570-1,580 |
| Ornith 1.5 | APEX-I-COMPACT | NVIDIA: RTX 5090 (32 GB), Xeon Platinum 8470Q (25 vCPU), 120 GB RAM | 316-319 | 6,830-6,860 |
| Ornith 1.5 | APEX-I-MINI | NVIDIA: RTX 5090 (32 GB), Xeon Platinum 8470Q (25 vCPU), 120 GB RAM | 289-300 | 6,300-6,380 |
| Ornith 1.5 | APEX-I-QUALITY | NVIDIA: RTX 5090 (32 GB), Xeon Platinum 8470Q (25 vCPU), 120 GB RAM | 289-298 | 6,040-6,390 |
| Gemma 4 26B-A4B | UD-Q3_K_M | NVIDIA: RTX 4070 Laptop (8 GB), Core i7-14700HX, 32 GB RAM | 92-93 | 2,240 |
| Gemma 4 26B-A4B | UD-Q2_K_XL | NVIDIA: RTX 4070 Laptop (8 GB), Core i7-14700HX, 32 GB RAM | 101-104 | 2,150-2,160 |
| Gemma 4 26B-A4B | UD-Q4_K_M | NVIDIA: RTX 4070 Laptop (8 GB), Core i7-14700HX, 32 GB RAM | 85 | 2,150 |
| Gemma 4 26B-A4B | UD-Q3_K_M | NVIDIA: RTX 5090 (32 GB), Xeon Platinum 8470Q (25 vCPU), 120 GB RAM | 353-356 | 7,880-8,230 |
| Gemma 4 26B-A4B | UD-Q2_K_XL | NVIDIA: RTX 5090 (32 GB), Xeon Platinum 8470Q (25 vCPU), 120 GB RAM | 327-340 | 7,380-8,000 |
| Gemma 4 26B-A4B | UD-Q4_K_M | NVIDIA: RTX 5090 (32 GB), Xeon Platinum 8470Q (25 vCPU), 120 GB RAM | 335-346 | 7,040-8,050 |
| Tiel Coder 35B-A3B | UD-Q3_K_XL | NVIDIA: RTX 4070 Laptop (8 GB), Core i7-14700HX, 32 GB RAM | 83 | 1,700 |
| Tiel Coder 35B-A3B | UD-IQ3_XXS | NVIDIA: RTX 4070 Laptop (8 GB), Core i7-14700HX, 32 GB RAM | 96-97 | 1,610 |
| Tiel Coder 35B-A3B | UD-Q4_K_XL | NVIDIA: RTX 4070 Laptop (8 GB), Core i7-14700HX, 32 GB RAM | 74-79 | 1,660-1,670 |
| Tiel Coder 35B-A3B | UD-Q3_K_XL | NVIDIA: RTX 5090 (32 GB), Xeon Platinum 8470Q (25 vCPU), 120 GB RAM | 337-345 | 6,570-6,830 |
| Tiel Coder 35B-A3B | UD-IQ3_XXS | NVIDIA: RTX 5090 (32 GB), Xeon Platinum 8470Q (25 vCPU), 120 GB RAM | 339-351 | 6,090-6,520 |
| Tiel Coder 35B-A3B | UD-Q4_K_XL | NVIDIA: RTX 5090 (32 GB), Xeon Platinum 8470Q (25 vCPU), 120 GB RAM | 316-322 | 6,750-6,870 |
Community reports
See community_reports.csv, and the community page for how to add yours.