Gemma 4 26B A4B (it)
Files and memory at 8K context
| Quant | File | Files | + KV + runtime | Smallest fit | Source |
|---|---|---|---|---|---|
| Q4_K_M lmstudio-community/gemma-4-26B-A4B-it-GGUF | 15.64 GiB | 1 | unknown | 20 GB GPU (cache unknown) | HF files |
| Q8_0 unsloth/gemma-4-26B-A4B-it-GGUF | 25.02 GiB | 1 | unknown | 32 GB GPU (cache unknown) | HF files |
| MLX 4-bit mixed precision mlx-community/gemma-4-26B-A4B-it-4bit | 14.29 GiB | 3 | unknown | 24 GB Mac (cache unknown) | HF files |
| MLX 8-bit mixed precision mlx-community/gemma-4-26B-A4B-it-8bit | 26.03 GiB | 6 | unknown | 48 GB Mac (cache unknown) | HF files |
No verified file: Q5_K_M, FP16. Sizes are exact file bytes (text weights; projectors listed separately). “+” = recurrent state or other terms unknown. Smallest fit adds 0.5 GiB spare; Macs use 75% of memory.
Quality
Artificial Analysis Intelligence Index
16.7 AA estimate
Variant: Gemma 4 26B A4B (Reasoning) · observed 2026-10-17 · source
Rank #39 of 75 researched 2026 models.
Arena text rating
1,434
95% CI 1427–1442 · 6,094 votes · published 2026-10-15 · dataset
Exact listed model name; hosted reasoning/quantization may differ from local checkpoint.
Scores describe the evaluated hosted variant, not each quantized file. AA and Arena are never averaged.
Reported speeds 12
| Hardware | Quant | Backend | Prompt t/s | Gen t/s | Date | Source |
|---|---|---|---|---|---|---|
| M4 Max (40c GPU) 128 GB | MLX 4bit | MLX | 1,391 | 96.1 | 2026-10-16 | omlx.ai |
| M5 Pro (20c GPU) 64 GB | MLX 4bit | MLX | 2,114 | 79.5 | 2026-10-15 | omlx.ai |
| M1 Max (24c GPU) 64 GB | MLX 4bit | MLX | 341 | 36 | 2026-10-13 | omlx.ai |
| M3 Max (40c GPU) 128 GB | MLX 4bit | MLX | 1,109 | 78.6 | 2026-10-11 | omlx.ai |
| M2 Ultra (60c GPU) 64 GB | MLX 4bit | MLX | 1,064 | 78.9 | 2026-10-07 | omlx.ai |
| NVIDIA GeForce RTX 3090 24 GB | Q4_K (subvariant not stated) | llama.cpp CUDA | 3,626 | 119 | unknown | hardware-corner.net |
| NVIDIA GeForce RTX 3090 24 GB | Q4_K (subvariant not stated) | llama.cpp CUDA | 2,453 | 108 | unknown | hardware-corner.net |
| NVIDIA GeForce RTX 3090 24 GB | Q4_K (subvariant not stated) | llama.cpp CUDA | 1,147 | 83 | unknown | hardware-corner.net |
| NVIDIA GeForce RTX 5090 32 GB | Q4_K (subvariant not stated) | llama.cpp CUDA | 8,799 | 180 | unknown | hardware-corner.net |
| NVIDIA GeForce RTX 5090 32 GB | Q4_K (subvariant not stated) | llama.cpp CUDA | 2,839 | 130 | unknown | hardware-corner.net |
| NVIDIA RTX PRO 6000 Blackwell 96 GB | Q4_K (subvariant not stated) | llama.cpp CUDA | 9,437 | 197 | unknown | hardware-corner.net |
| Surface Laptop Ultra,128GB,RTX Spark | UD-Q4_K_M + MTP | llama.cpp | — | 101 | unknown | source |
Reported by the linked sources, not reproduced by LocalFit. All rows in the speed database
Community threads
- Reliable tool calling for Markdown vaults and LLM wikis w/ Gemma Hugging Face · 2026-09-23
- Chat template may re-inject prior-turn reasoning during multi-turn tool use → repetition loops Hugging Face · 2026-06-10
Qualitative signals only; titles are not verified claims.