There is no established overall winner among MLX’s 4-bit formats. The current mlx-lm converter defaults to affine quantization, but it also offers MXFP4 and NVFP4. The documented defaults tell you how conversion starts—not which format gives the best quality, speed, or memory use for your model and workload.
What is the best 4-bit format in MLX?
It depends on what you are optimizing and what model, runtime, and Apple silicon device you use. The available official sources describe converter options and examples, but do not provide a controlled, head-to-head benchmark establishing that affine, MXFP4, or NVFP4 is best overall.
As an Amazon Associate I earn from qualifying purchases.
That means the converter’s default is not evidence that affine is inferior—or that another mode is superior. To make a defensible choice, compare formats under the same conditions and measure the outcomes that matter to you.
What the MLX converter defaults to
The mlx-lm conversion CLI accepts --q-mode values affine, mxfp4, nvfp4, and mxfp8. The default mode is affine. You can also set --q-bits and --q-group-size; the implementation’s defaults are:
#1 Best Overall
| Mode | Default bit width | Default group size | 4-bit option? |
|---|---|---|---|
| Affine | 4 | 64 | Yes |
| MXFP4 | 4 | 32 | Yes |
| NVFP4 | 4 | 16 | Yes |
| MXFP8 | 8 | 32 | No |
These are defaults in the converter implementation on the repository’s moving main branch, not guarantees for every released version. Check the implementation for the version you are using: mlx-lm converter source.
How to choose a mode fairly
First decide what “best” means for your use. A format that preserves more quality may not produce the smallest artifact or the fastest generation on your particular setup. Compare modes using the same model and device, and hold the runtime and workload constant.
- Output quality: Use a named evaluation set or repeatable task-specific prompts, and state the scoring method. For a practical comparison, include representative prompts from the tasks you actually care about.
- Artifact size and memory: Record the converted model’s file size and, where relevant, peak memory during loading and inference. Account for quantization metadata, not just nominal bit width.
- Speed: Keep the runtime, prompt length, generated-token count, and device the same. Report the metric you measure, such as generation speed, rather than calling one mode faster without a test.
- Compatibility: Verify that the model architecture and runtime support the chosen mode, and confirm that the converted model loads and runs as intended.
- Bit allocation: Distinguish a uniform 4-bit conversion from a mixed-bit recipe. A model described as “4-bit” may have selected layers stored at a different precision.
For reproducibility, identify the model, mlx-lm version or commit, Apple silicon device, conversion settings, workload, and evaluation metrics. Without those details, a comparison may not apply to another setup.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Why “4-bit” does not equal four stored bits per weight
Nominal bit width does not include every bit needed to store quantized weights. Scales, biases, and other metadata add overhead. The Hugging Face transformers-to-MLX guide estimates that 4-bit quantization with group size 64 typically uses about 4.5 effective bits per weight after scale and bias metadata. That is the guide’s rough estimate, not a universal measurement for every model or quantization mode.
Rank #3
- NVIDIA Volta GV100 Architecture — 4,608 CUDA Cores, 640 1st-Gen Tensor Cores delivering 14 TFLOPS FP32 and 112 TFLOPS deep learning performance for AI training, inference, HPC, and scientific computing workloads
- 32GB HBM2 ECC Memory — 900 GB/s Bandwidth — High-bandwidth memory on a 4096-bit bus with ECC error correction provides the memory capacity and throughput required for the largest AI models, simulations, and datasets
- PCIe 3.0 x16 Interface — 250W TDP — Standard PCIe Gen3 connectivity with passive cooling designed for enterprise rack server deployment in HPE ProLiant, Dell PowerEdge, and Supermicro platforms with adequate chassis airflow
- NVLink — Scale to 96GB Unified Memory — Connect two V100 GPUs via NVLink at 300 GB/s bi-directional bandwidth to scale GPU memory from 32GB to 96GB for larger AI training and HPC workloads
- Multi-Precision Computing — Supports FP64 (7 TFLOPS), FP32 (14 TFLOPS), FP16 (112 TFLOPS) and INT8 precision modes for flexible deployment across training, inference, and scientific simulation workloads
When mixed-bit quantization may be worth testing
Uniform 4-bit conversion is not the only option. The converter implementation includes mixed-bit recipes that assign more bits to selected components—including value projections, down projections, and the language-model head—while using fewer bits elsewhere. Apple’s WWDC25 demonstration also shows a custom predicate assigning 6 bits to lm_head and embed_tokens, 4 bits to other quantizable layers, and skipping modules that cannot be quantized.
This flexibility can make mixed precision a useful candidate when you want to explore a different balance between quality and storage. It does not, by itself, establish that mixed precision will improve a particular model’s results. Apple’s example and converter details are available in the WWDC25 MLX LM session and the converter implementation.
Rank #4
- Durable: Constructed with high-quality metal, this mining frame ensures long-lasting durability and full protection for your GPU mining rig and electronic devices.
- Efficient Cooling: Designed for enhanced air convection, this mining case maximizes heat dissipation, helping to extend the service life of your GPUs during intensive mining operations.
- Professional Build: Features non-slip rubber feet and EVA foam on the crossbar to prevent damage to your graphic cards. Perfect for securing and protecting your GPUs in a mining rig setup.
- Stackable Design: This mining frame supports stackable configurations, allowing you to expand your GPU mining setup easily with additional mining cases or stacking brackets (sold separately).
- Stable and Secure: Equipped with rubber feet, this mining case prevents shaking and moving, keeping your mining rig stable during operation.
Practical takeaway for an MLX conversion
Start with the mode that matches your goal, not the assumption that the default is designed to lose. The converter’s defaults are a starting configuration; choosing a winner requires a workload-specific comparison of quality, storage and memory, speed, and compatibility. If you do not have those measurements, describe the choice as an experiment or preference rather than a proven best format.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
