Qwen3.5-9B-Instruct-MXFP4-GGUF
MXFP4 (OCP Microscaling FP4) GGUF quantization of the Qwen3.5-9B-Instruct multimodal language model with thinking disabled by default.
About MXFP4
MXFP4 (E2M1) is the open-standard 4-bit floating-point format defined by the OCP Microscaling Formats (MX) specification. Unlike block-scaled integer formats (Q4_K_M), MXFP4 uses per-group scaling factors (block size 32) with shared exponent via the E2M1 format (2 exponent bits, 1 mantissa bit). This provides better dynamic range utilization than integer formats and works across GPU vendors (NVIDIA, AMD, Intel) and CPUs.
Key Advantages
- Cross-vendor: Runs on any GPU or CPU (no vendor lock-in)
- E2M1 format: 2 exponent bits + 1 mantissa bit + shared scale per 32 elements
- Floating-point: Better dynamic range than integer quantization
- Open standard: OCP MX specification, widely supported
Repo Contents
| Filename | Type | Size | Description |
|---|---|---|---|
qwen35-9b-instruct-mxfp4.gguf | Text model | 5.31 GB | MXFP4 quantized model (no MTP head) |
mmproj-qwen35-9b-f16.gguf | Vision encoder | 922 MB | SigLIP vision projector (F16) |
Quantization Details
| Property | Value |
|---|---|
| Format | MXFP4 (OCP E2M1) |
| Block size | 32 elements |
| BPW | ~4.74 |
| Architecture | qwen35 (no MTP, 427 tensors) |
| Target hardware | Universal (CPU, AMD, NVIDIA, Intel) |
| Thinking | Disabled by default (opt-in via enable_thinking=true) |
| MTP | Disabled (--no-mtp) |
Usage
llama.cpp
# Basic text generation (thinking disabled by default)
./llama-cli -m qwen35-9b-instruct-mxfp4.gguf -p "What is the capital of France?" -n 256
# Vision (requires mmproj)
./llama-cli -m qwen35-9b-instruct-mxfp4.gguf --mmproj mmproj-qwen35-9b-f16.gguf -p "Describe this image" --image photo.jpg -n 256
# Enable thinking
./llama-cli -m qwen35-9b-instruct-mxfp4.gguf -p "Solve: 2+2=?" -n 512 -e enable_thinking=true
Python
from llama_cpp import Llama
llm = Llama(
model_path="qwen35-9b-instruct-mxfp4.gguf",
n_ctx=8192,
chat_format="qwen3",
)
# Basic chat (no thinking)
output = llm.create_chat_completion([{"role": "user", "content": "What is AI?"}])
# With thinking enabled
output = llm.create_chat_completion(
[{"role": "user", "content": "Solve 2+2=?"}],
extra_body={"enable_thinking": True},
)
Download
huggingface-cli download FreedomAISVR/Qwen3.5-9B-Instruct-MXFP4-GGUF --local-dir . --local-dir-use-symlinks False
Original Model
Qwen3.5-9B is Alibaba Cloud's efficient multimodal foundation model (Apache 2.0, March 2026) featuring:
- Hybrid Gated DeltaNet + Gated Attention (3:1 ratio)
- 262K native context (extensible to 1M)
- Text + Image + Video input
- 201 language support
Conversion
python convert_hf_to_gguf.py --no-mtp --outfile qwen35-9b-f16.gguf D:\qwen35-9b-src
python convert_hf_to_gguf.py --mmproj --outfile mmproj-qwen35-9b-f16.gguf D:\qwen35-9b-src
llama-quantize.exe qwen35-9b-f16.gguf qwen35-9b-instruct-mxfp4.gguf MXFP4
License
Apache 2.0 (same as Qwen3.5-9B)