Base model: Qwen/Qwen3.5-4B-Base (4.21B text backbone)
Trained by: Michael Anthony Falabella
Method: LoRA fine-tune (bf16, rank 16), merged into the base model
Format: All files below are GGUF (llama.cpp / Ollama / LM Studio compatible)

Available models (quantizations)

Quant	File	Size	Quality vs. speed
F16	`qwen-3.5-4b-Coyote.F16.gguf`	7.85 GB	Maximum quality, largest/slowest
Q8_0	`qwen-3.5-4b-Coyote.Q8_0.gguf`	4.17 GB	Near-F16 quality, recommended default
Q6_K	`qwen-3.5-4b-Coyote.Q6_K.gguf`	3.23 GB	Great balance
Q4_K_M	`qwen-3.5-4b-Coyote.Q4_K_M.gguf`	2.52 GB	Smallest, fastest, best for low VRAM

Architecture qwen35, 32 transformer blocks, 426 tensors, 248,320-token vocabulary.

Quick start

ollama create qwen3.5-4b-coyote -f Modelfile

with a Modelfile containing:

FROM ./qwen-3.5-4b-Coyote.Q8_0.gguf

Then:

ollama run qwen3.5-4b-coyote

./llama-cli -m qwen-3.5-4b-Coyote.Q8_0.gguf -p "Write a Python function that ..." -cnv

Evaluated on the BenchLocal coding benchpacks (as of the latest release):

Pack	Score
ToolCall-15	67%
BugFind-15	65%

Apache 2.0 — inherits the base model license. See Qwen/Qwen3.5-4B-Base for the underlying model license and terms.