Qwen3.6-27B
Files and memory at 8K context
| Quant | File | Files | + KV + runtime | Smallest fit | Source |
|---|---|---|---|---|---|
| Q4_K_M unsloth/Qwen3.6-27B-GGUF | 15.66 GiB | 1 | 17.2 GiB+ | 20 GB GPU | HF files |
| Q5_K_M unsloth/Qwen3.6-27B-GGUF | 18.17 GiB | 1 | 19.7 GiB+ | 24 GB GPU | HF files |
| Q8_0 unsloth/Qwen3.6-27B-GGUF | 26.63 GiB | 1 | 28.1 GiB+ | 32 GB GPU | HF files |
| MLX 4-bit mlx-community/Qwen3.6-27B-4bit | 14.95 GiB | 3 | 16.5 GiB+ | 24 GB Mac | HF files |
| MLX 8-bit mlx-community/Qwen3.6-27B-8bit | 27.48 GiB | 6 | 29.0 GiB+ | 48 GB Mac | HF files |
No verified file: 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
Unknown
Arena text rating
Unknown
Scores describe the evaluated hosted variant, not each quantized file. AA and Arena are never averaged.
Reported speeds 19
| Hardware | Quant | Backend | Prompt t/s | Gen t/s | Date | Source |
|---|---|---|---|---|---|---|
| M4 Max (40c GPU) 64 GB | MLX 4bit | MLX | 233 | 26.8 | 2026-10-06 | omlx.ai |
| M5 Max (32c GPU) 36 GB | MLX 4bit | MLX | 750 | 24.8 | 2026-09-17 | omlx.ai |
| M5 Pro (20c GPU) 48 GB | MLX 4bit | MLX | 478 | 17.1 | 2026-08-25 | omlx.ai |
| M1 Max (32c GPU) 64 GB | MLX 4bit | MLX | 90.1 | 18.3 | 2026-08-21 | omlx.ai |
| M4 Max (32c GPU) 36 GB | MLX 4bit | MLX | 209 | 21.6 | 2026-08-15 | omlx.ai |
| M3 Ultra (80c GPU) 512 GB | MLX 4bit | MLX | 442 | 33.8 | 2026-08-14 | omlx.ai |
| M5 Max (40c GPU) 128 GB | MLX 4bit | MLX | 945 | 32.1 | 2026-08-13 | omlx.ai |
| M3 Ultra (60c GPU) 96 GB | MLX 4bit | MLX | 327 | 36.5 | 2026-08-11 | omlx.ai |
| M4 Pro (16c GPU) 48 GB | MLX 4bit | MLX | 105 | 14.8 | 2026-08-10 | omlx.ai |
| M2 Max (30c GPU) 64 GB | MLX 4bit | MLX | 97 | 19.2 | 2026-08-07 | omlx.ai |
| M2 Max (38c GPU) 32 GB | MLX 4bit | MLX | 122 | 20 | 2026-08-05 | omlx.ai |
| M4 Pro (20c GPU) 48 GB | MLX 4bit | MLX | 131 | 14.9 | 2026-08-04 | omlx.ai |
| NVIDIA GeForce RTX 5090 32 GB | Q4_K_M | llama.cpp CUDA | 3,874 | 75.6 | unknown | hardware-corner.net |
| NVIDIA GeForce RTX 5090 32 GB | Q4_K_M | llama.cpp CUDA | 3,486 | 72.8 | unknown | hardware-corner.net |
| NVIDIA GeForce RTX 5090 32 GB | Q4_K_M | llama.cpp CUDA | 2,941 | 69.5 | unknown | hardware-corner.net |
| NVIDIA GeForce RTX 5090 32 GB | Q4_K_M | llama.cpp CUDA | 1,957 | 63.7 | unknown | hardware-corner.net |
| NVIDIA GeForce RTX 5090 32 GB | Q4_K_M | llama.cpp CUDA | 1,180 | 54.6 | unknown | hardware-corner.net |
| Apple Mac Studio M3 Ultra, 256 GB unified memory (GPU core count not stated) | Q4_K_M | Ollama (llama.cpp Metal) | 95 | 28.6 | 2026-08-28 | terminalbytes.com |
| Apple M5 Pro, 128 GB RAM (GPU core count not stated) | Q4_K_M | llama.cpp Metal (llama-server via brew) | — | 25.6 | 2026-04-22 | simonwillison.net |
Reported by the linked sources, not reproduced by LocalFit. All rows in the speed database
Community threads
- GGUF available — Cerebellum v4 (ablation-guided mixed-precision) Hugging Face · 2026-05-01
- Qwen3.6 27b has template issue in ralph loop on opencode: `No user query found in messages` Hugging Face · 2026-04-24
Qualitative signals only; titles are not verified claims.