Qwen/Qwen3.5-4B
Full fine-tune vs Qwen/Qwen3.5-4B-Base. Fine-tune: 93% of the weights changed, across 32 of 32 layers.
Tensors changed: Embeddings 3/5, MLP 103/199, Attention 205/324, Norms 113/209, other 1/1
583 derivatives checked
- unsloth/Qwen3.5-4B-GGUF: Same weights, converted
- cyankiwi/Qwen3.5-4B-AWQ-4bit: Quantization
- QuantTrio/Qwen3.5-4B-AWQ: Quantization
- lmstudio-community/Qwen3.5-4B-GGUF: Quantization
- bartowski/Qwen_Qwen3.5-4B-GGUF: Same weights, converted
- unsloth/Qwen3.5-4B: Identical copy
- unsloth/Qwen3.5-4B-MTP-GGUF: Same weights, converted
- empero-ai/Qwen3.8-4B-Distill: Full fine-tune
- Jackrong/Qwen3.5-4B-Claude-4.6-Opus-Reasoning-Distilled-GGUF: Full fine-tune
- mlx-community/Qwen3.5-4B-MLX-4bit: Quantization
- jica98/qwen3.5-4B-super-coder: Adapter
- numind/NuExtract3: Full fine-tune
- athrael-soju/colqwen3.5-4.5B-v3: LoRA-style
- TokenRhythm/NeoHorse-1-4B: Full fine-tune (fewer layers)
- RedHatAI/Qwen3.5-4B-FP8-dynamic: Quantization
- sky7350/Mica-v0.1-4B: LoRA-style (fewer layers)
- alibiserikbay/JevK5: LoRA-style (fewer layers)
- mlx-community/Qwen3.5-4B-OptiQ-4bit: Quantization
- org2ai/Wald-4B: Partial fine-tune
- numind/NuExtract3-FP8: Quantization
- lmstudio-community/Qwen3.5-4B-MLX-4bit: Quantization
- Jackrong/Qwen3.5-4B-Claude-4.6-Opus-Reasoning-Distilled-v2-GGUF: Full fine-tune
- huihui-ai/Huihui-Qwen3.5-4B-abliterated: LoRA-style
- Jackrong/Qwopus3.5-4B-Coder-MTP-GGUF: LoRA-style
- z-lab/Qwen3.5-4B-DFlash: Different architecture
- insraq/Qwen3.5-4B-EmperoAI-Qwen3.8-Distill-Heretic-Abliterated-MTP-GGUF: Full fine-tune
- RedHatAI/Qwen3.5-4B-quantized.w4a16: Quantization
- AtomicChat/Qwen3.5-4B-GGUF: Quantization
- Qwen/Qwen-Drive-1.0-4B: Full fine-tune
- lmstudio-community/Qwen3.5-4B-MLX-8bit: Quantization
76 other repos hold byte-for-byte identical weight files.