Qwen3.5 4B (Q3_K_M)on CPU Only

Q3_K_M CPU Only

Overview

Qwen3.5 4B is a 4.66B parameter dense language model by Qwen, with code, multilingual, thinking, tool-calls, vision capabilities. It supports a context window of up to 262,144 tokens.

Qwen3.5 4B is a model from Alibaba's Qwen 3.5 family built on the Gated Delta Networks hybrid architecture with 4.66 billion parameters, widely regarded as the community sweet spot for performance per watt. It is natively multimodal, processing text, images, and video, with built-in thinking capabilities for chain-of-thought reasoning. The model supports a 262K context window and covers over 201 languages, nearly matching previous-generation 80B MoE models on coding benchmarks. Released under the Apache 2.0 license, it runs in roughly 3 GB of VRAM at Q4, delivering fast and stable self-hosted deployment on consumer hardware.

At Q3_K_M quantization (low quality tier), the model weighs 2.14 GB. This exceeds the 0 GB of VRAM on CPU Only. Inference is still possible via CPU offload or memory-mapped loading from disk, but expect significantly reduced performance.

A CPU-only configuration with no GPU acceleration. Inference runs entirely on the CPU, which is significantly slower than GPU-accelerated setups but requires no special hardware. Performance and maximum model size depend on available system RAM. Suitable for testing, development, or deployments where no GPU is available.

Hardware Requirements

Model size	2.14 GB
VRAM available	0 GB
VRAM used	0 GB
Min RAM required	2.1 GB
GPU layers	0 / 32
Context size	262,144
Backend	cpu
Flash attention	No

Performance Notes

Reduced performance

32 of 32 layers run on CPU. Your system needs at least 2.1 GB RAM. If system RAM is insufficient, data will be read from disk on demand, with significant performance impact.

Deploy

Server Local

Command

helmfile --state-values-file <(curl -s https://www.prositronic.eu/values/qwen3-5-4b/q3_k_m/cpu.yaml) apply

Generated values.yaml

/values/qwen3-5-4b/q3_k_m/cpu.yaml

Loading values…

Install llama.cpp

Install llama.cpp from the official build instructions for your platform and backend.

Download Model

curl -L -o qwen3-5-4b.gguf "https://huggingface.co/unsloth/Qwen3.5-4B-GGUF/resolve/main/Qwen3.5-4B-Q3_K_M.gguf"

Start Server

llama-server \
  -m qwen3-5-4b.gguf \
  --n-gpu-layers 0 \
  --ctx-size 262144

Verify

curl http://localhost:8080/health

Explore

• Configure a deployment
• Browse models
• Browse hardware

Other quantizations

• Qwen3.5 4B (Q8_0)
• Qwen3.5 4B (Q5_K_M)
• Qwen3.5 4B (Q4_K_M)

Other hardware

• Qwen3.5 4B on NVIDIA A100 80GB
• Qwen3.5 4B on NVIDIA H100
• Qwen3.5 4B on NVIDIA RTX 3090

Frequently Asked Questions

How much VRAM does Qwen3.5 4B (Q3_K_M) need?

The Q3_K_M quantization of Qwen3.5 4B requires 2.14 GB. The 0 GB of VRAM on CPU Only is insufficient for GPU layers, so inference runs on CPU.

Can I run Qwen3.5 4B on CPU Only?

It is possible but not recommended. CPU Only does not have enough VRAM to accelerate Qwen3.5 4B (Q3_K_M), so inference will rely on CPU and system RAM.

What is quantization?

Quantization reduces a model's numerical precision from its original floating-point format to a more compact representation. This shrinks the file size and VRAM footprint, making it possible to run large models on consumer hardware. The trade-off is a small reduction in output quality. Q3_K_M compresses Qwen3.5 4B from its original size down to 2.14 GB.

What quantization should I choose for Qwen3.5 4B?

Q3_K_M is a low-quality quantization. Higher-quality quants (Q8, Q6) preserve more model accuracy but need more VRAM. Lower quants (Q4, Q3, Q2) reduce VRAM usage at the cost of some quality. Choose based on your available hardware and quality requirements.

Why are some layers offloaded to CPU?

CPU Only has 0 GB of VRAM, but Qwen3.5 4B (Q3_K_M) requires approximately 2.14 GB. Only 0 of 32 layers fit in VRAM; the remaining layers run on CPU, which is slower but still functional.

How do I run Qwen3.5 4B (Q3_K_M) with Ollama?

Run ollama run qwen3.5:4b-q3_k_m to start Qwen3.5 4B (Q3_K_M). Ollama handles downloading the model weights automatically on first run.

Last updated: March 13, 2026