Qwen3.5-4B-Instruct-Revised-GGUF

library name: transformers license: apache 2.0 pipeline tag: image text to text base model: Qwen/Qwen3.5 4B base model relation: quantized tags: RevisedGGUF Qwen3.5 4B | Revised GGUF Quantizations Revised GGUF quantizations of the official Qwen 3.5 4B model, converted locally using llama.cpp b9009. These modified GGUFs have chat template fixes/improvements while maintaning original pure functionality. They're nicknamed Revised as it is a change to the original source. This model is no different capability wise from the Pure GGUF variant. Think of it as an improved version of it. This version comes with a baked in fixed chat template made by froggeric. Drop in Jinja templates that fix rendering errors, token waste, and missing features in the official Qwen chat templates. Works in LM Studio, llama.cpp, vLLM, MLX, oMLX, and any engine that supports HuggingFace Jinja templates. — froggeric Why was this variant made? The official Qwen templates have bugs and missing features, such as not having a clean toggle for thinking, empty thinking blocks spam context, developer role rejected, and more. By using this version, you get the original functionality along with improvements. I recommend using this model over the Pure GGUF, as it's the same thing except a small chat template change If you're wanting the original Pure version, check out Qwen3.5 4B Instruct Pure GGUF Picking a Quantization | Quant | VRAM | Notes | | | | | | q8 0 and above | 16GB+ | Closest to original quality | | q5 k m | 12GB+ | Slightly higher quality output | | q4 k m | 8–12GB | Best balance for most users | Example: q4 k m at a 64k context uses 7GB VRAM . Note: Download mmproj f16.gguf alongside the main model file to enable vision capabilities. Controlling Thinking Mode This variant includes a way of controlling thinking via the system prompt. By default, this model is in thinking mode. To toggle thinking, you need to drop or anywhere in your system prompt. An example of this in a system prompt would be: Quantization Process 1. Weights pulled directly from Qwen/Qwen3.5 4B on HuggingFace. 2. Skipped: .git , .gitattributes , LICENSE , README . 3. Edit chat template.jinja and tokenizer config.json 4. Converted to F16 GGUF via convert hf to gguf.py (llama.cpp). 5. Quantized locally with llama quantize . 6. Uploaded to HuggingFace. Official Model Card