LocalFit

Qwen3.5-4B

Qwen/Qwen3.5-4B

Released2026-03-02Size4B denseArchitecturehybrid linear attnLicenseapache-2.0AA rank#48 of 75

Files and memory at 8K context

QuantFileFiles+ KV + runtimeSmallest fitSource
Q4_K_M
unsloth/Qwen3.5-4B-GGUF
2.55 GiB13.8 GiB+8 GB GPUHF files
Q5_K_M
unsloth/Qwen3.5-4B-GGUF
2.93 GiB14.2 GiB+8 GB GPUHF files
Q8_0
unsloth/Qwen3.5-4B-GGUF
4.17 GiB15.4 GiB+8 GB GPUHF files
MLX 4-bit
mlx-community/Qwen3.5-4B-4bit
2.83 GiB14.1 GiB+8 GB MacHF files
MLX 8-bit
mlx-community/Qwen3.5-4B-8bit
4.78 GiB16.0 GiB+16 GB MacHF 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

13.1 AA estimate

Variant: Qwen3.5 4B (Reasoning) · observed 2026-10-17 · source

Rank #48 of 75 researched 2026 models.

Arena text rating

Unknown

Scores describe the evaluated hosted variant, not each quantized file. AA and Arena are never averaged.

Reported speeds 6

HardwareQuantBackendPrompt t/sGen t/sDateSource
MacBook Air M5, 32 GB (10 CPU / 10 GPU cores)Q4_K_Mllama.cpp Metal63129.42026-04-06r/LocalLLaMA
Steam Deck (AMD Custom GPU 0932 / RADV VANGOGH, Vulkan)Q4_K_Mllama.cpp Vulkan18922.62026-08-20llama.cpp #10879
Steam Deck (AMD Custom GPU 0932 / RADV VANGOGH, Vulkan)Q4_K_Mllama.cpp Vulkan19022.72026-08-20llama.cpp #10879
NVIDIA GeForce RTX 3090 24 GBQ4_K_XL (unsloth UD)llama.cpp Vulkan5,2341632026-04-23r/LocalLLaMA
Intel Arc Pro B70 32 GBQ4_K_XL (unsloth UD)llama.cpp Vulkan1,43160.52026-04-23r/LocalLLaMA
Intel Arc Pro B70 32 GBQ4_K_XL (unsloth UD)llama.cpp SYCL1,76779.12026-04-23r/LocalLLaMA

Reported by the linked sources, not reproduced by LocalFit. All rows in the speed database

Community threads

Qualitative signals only; titles are not verified claims.