byteshape/Qwen3-4B-Instruct-2507-GGUF

Qwen3-4B-Instruct-2507 GGUF (ShapeLearn Quantized)

This is a GGUF-quantized version of Qwen3-4B-Instruct-2507 produced with ByteShape's ShapeLearn, which learns the optimal datatype per tensor to maintain high quality even at very low bit lengths (the exclusive focus of this release).

To learn more about ShapeLearn and to see detailed benchmarks across GPUs, CPUs, and even the Raspberry Pi, please visit our blog.

If you have questions or want to share feedback, reach us on Reddit.

How to Pick a Model

We provide CPU and GPU optimized variants for llama.cpp:

• CPUs: KQ quantization is preferred due to GGML kernel efficiency.
• Nvidia GPUs: IQ quantization delivers faster throughput on modern architectures.

Each hardware target includes a range of models covering different size–quality tradeoffs.

The charts below show quality vs. tokens per second for each device, comparing ShapeLearn models with Unsloth baselines.

Selection rule: Choose the model with the highest quality at your target throughput or the fastest model that still meets your required quality.

GGUF-KQ Models: Best for CPU

Table sorted by inference speed (match the chart’s numbers to model IDs):

Model ID	Bits/Weight	Model Size	Normalized Quality
KQ-1	2.77	1.4 GB	70.33%
KQ-2	2.95	1.49 GB	79.81%
KQ-3	3.19	1.61 GB	87.31%
KQ-4	3.34	1.69 GB	92.04%
KQ-5	3.45	1.74 GB	93.01%
KQ-6	3.66	1.84 GB	94.46%
KQ-7	3.87	1.95 GB	95.89%
KQ-8	4.31	2.17 GB	98.44%
KQ-9	4.74	2.39 GB	98.95%

GGUF-IQ Models: Best for GPU

Table sorted by inference speed (match the chart’s numbers to model IDs):

Model ID	Bits/Weight	Model Size	Normalized Score
IQ-1	2.55	1.29 GB	69.87%
IQ-2	2.76	1.39 GB	83.32%
IQ-3	2.94	1.49 GB	89.04%
IQ-4	3.07	1.55 GB	92.32%
IQ-5	3.31	1.67 GB	92.45%
IQ-6	3.55	1.79 GB	95.16%
IQ-7	4.04	2.04 GB	99.25%
IQ-8	4.54	2.29 GB	99.80%