Qwen3.8-27B
Files and memory at 8K context
| Quant | File | Files | + KV + runtime | Smallest fit | Source |
|---|---|---|---|---|---|
| Q4_K_M bartowski/Qwen3.8-27B-GGUF | 16.24 GiB | 1 | 17.7 GiB+ | 20 GB GPU | HF files |
| Q5_K_M bartowski/Qwen3.8-27B-GGUF | 19.49 GiB | 1 | 21.0 GiB+ | 24 GB GPU | HF files |
| Q8_0 bartowski/Qwen3.8-27B-GGUF | 27.12 GiB | 1 | 28.6 GiB+ | 32 GB GPU | HF files |
| MLX 4-bit mlx-community/Qwen3.8-27B-4bit | 14.95 GiB | 3 | 16.5 GiB+ | 24 GB Mac | HF files |
| MLX 8-bit mlx-community/Qwen3.8-27B-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
33.7
Variant: Qwen3.8 27B (Xhigh) · observed 2026-10-17 · source
Rank #13 of 75 researched 2026 models.
Arena text rating
1,441
95% CI 1435–1446 · 23,294 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 53
| Hardware | Quant | Backend | Prompt t/s | Gen t/s | Date | Source |
|---|---|---|---|---|---|---|
| M5 Max (40c GPU) 64 GB | MLX 4bit | MLX | 1,000 | 32.9 | 2026-10-16 | omlx.ai |
| M5 Pro (20c GPU) 64 GB | MLX 4bit | MLX | 516 | 17.7 | 2026-10-15 | omlx.ai |
| M4 Pro (20c GPU) 48 GB | MLX 4bit | MLX | 134 | 14.9 | 2026-10-15 | omlx.ai |
| M1 Max (32c GPU) 32 GB | MLX 4bit | MLX | 106 | 18.2 | 2026-10-15 | omlx.ai |
| M2 Max (38c GPU) 32 GB | MLX 4bit | MLX | 130 | 18.6 | 2026-10-14 | omlx.ai |
| M1 Max (24c GPU) 32 GB | MLX 4bit | MLX | 78.6 | 15.1 | 2026-10-13 | omlx.ai |
| M4 Pro (16c GPU) 48 GB | MLX 4bit | MLX | 75.5 | 9.8 | 2026-10-13 | omlx.ai |
| M3 Max (40c GPU) 128 GB | MLX 4bit | MLX | 233 | 21.9 | 2026-10-13 | omlx.ai |
| M3 Ultra (60c GPU) 256 GB | MLX 4bit | MLX | 353 | 37.8 | 2026-10-13 | omlx.ai |
| M4 Max (40c GPU) 128 GB | MLX 4bit | MLX | 268 | 27.8 | 2026-10-12 | omlx.ai |
| M2 Ultra (60c GPU) 64 GB | MLX 4bit | MLX | 204 | 32.1 | 2026-10-03 | omlx.ai |
| NVIDIA GeForce RTX 3090 24 GB | Q4_K_S | llama.cpp CUDA | 1,308 | 40.3 | unknown | hardware-corner.net |
| NVIDIA GeForce RTX 3090 24 GB | Q4_K_S | llama.cpp CUDA | 767 | 34 | unknown | hardware-corner.net |
| NVIDIA GeForce RTX 4090 24 GB | Q4_K_S | llama.cpp CUDA | 2,963 | 46.2 | unknown | hardware-corner.net |
| NVIDIA GeForce RTX 4090 24 GB | Q4_K_S | llama.cpp CUDA | 1,918 | 38.4 | unknown | hardware-corner.net |
| NVIDIA GeForce RTX 5090 32 GB | Q4_K_S | llama.cpp CUDA | 3,750 | 74.8 | unknown | hardware-corner.net |
| NVIDIA GeForce RTX 5090 32 GB | Q4_K_S | llama.cpp CUDA | 1,146 | 29 | unknown | hardware-corner.net |
| NVIDIA GeForce RTX 5090 32 GB | Q4_K_S | llama.cpp CUDA | 718 | 26.2 | unknown | hardware-corner.net |
| NVIDIA GeForce RTX 5090 32 GB | Q4_K_S | llama.cpp CUDA | 461 | 22.8 | unknown | hardware-corner.net |
| 2x NVIDIA GeForce RTX 5060 Ti 16 GB (15,849 MiB each, multi-GPU) multi-device | Q4_K_S | llama.cpp CUDA | 1,168 | 22.5 | unknown | hardware-corner.net |
| 2x NVIDIA GeForce RTX 5060 Ti 16 GB (15,849 MiB each, multi-GPU) multi-device | Q4_K_S | llama.cpp CUDA | 721 | 18.5 | unknown | hardware-corner.net |
| 2x NVIDIA GeForce RTX 5060 Ti 16 GB (15,849 MiB each, multi-GPU) multi-device | Q4_K_S | llama.cpp CUDA | 513 | 15.6 | unknown | hardware-corner.net |
| Apple M5 Max, 128 GB unified memory (GPU core count not stated) | Q4_K_S | llama.cpp (GPU backend not stated) | 834 | 31.4 | unknown | hardware-corner.net |
| Apple M5 Max, 128 GB unified memory (GPU core count not stated) | Q4_K_S | llama.cpp (GPU backend not stated) | 425 | 18.2 | unknown | hardware-corner.net |
| NVIDIA GB10 (Dell Pro Max, 128 GB unified memory) | Q4_K_S | llama.cpp CUDA | 813 | 12.2 | unknown | hardware-corner.net |
Reported by the linked sources, not reproduced by LocalFit. Showing 25 of 53. All rows in the speed database
Community threads
- Qwen3.8-27B-Flash-MTP-GGUF--5.79 GB. Code ~70-90 t/s, editing ~100-200 t/s. Hugging Face · 2026-10-16
- Intermittent Persian language-instruction failures after tool results — Alibaba via OpenRouter Hugging Face · 2026-10-12
- https://www.reddit.com/r/LocalLLaMA/comments/1wha6sv/i_benchmarked_ifmk2horizon7b_on_16gb_vram/ Reddit · 2026-09-15
- https://www.reddit.com/r/LocalLLaMA/comments/1x37q7b/update_qwen_38_27b_140_toks_on_single_rtx_3090/ Reddit · 2026-10-11
- https://www.reddit.com/r/LocalLLM/comments/1x70juc/qwen38_27b_took_7_hours_on_a_task_and_didnt_make/ Reddit · 2026-10-15
Qualitative signals only; titles are not verified claims.