Qwen3 is the latest generation of large language models in Qwen series, offering a comprehensive suite of dense and mixture-of-experts (MoE) models. Built upon extensive training, Qwen3 delivers groundbreaking advancements in reasoning, instruction-following, agent capabilities, and multilingual support, with the following key features: - Uniquely support of seamless switching between thinking mode (for complex logical reasoning, math, and coding) and non-thinking mode (for efficient, general-purpose dialogue) within single model, ensuring optimal performance across various scenarios. - Significantly enhancement in its reasoning capabilities, surpassing previous QwQ (in thinking mode) and…
Qwen3-32B-MXFP4-W4A4 is an open-weight model for text generation from Cezar, released under Apache License 2.0. It has 32.8B parameters and a 40,960-token context. At 16-bit it needs about 78.6 GB of GPU memory, which fits on 1x MI300X from $1.85 an hour, at the lowest prices in the SAVRN Index. It draws 8 downloads a month.
MXFP4 W4A4 quantization of (revision 9216db5781bf21249d130ec9da846c4624c16137, BF16).
Parameters32.8B
Context40,960
Weights19.7 GB
Licenseapache-2.0
AccessOpen weights
Monthly Downloads8
Runs On
What it takes to serve Qwen3-32B-MXFP4-W4A4 (32.8B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.
| Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits |
|---|---|---|---|---|---|
| 16-bit | 65.5 GB | 78.6 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 8-bit | 32.8 GB | 39.3 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 16.4 GB | 19.7 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Oct 9, 2026.
Qwen3-32B-MXFP4-W4A4 on every accelerator the SAVRN Index prices, at every precision
Model Card
By Cezar, published under apache-2.0, revision a49f042511ac.
Benchmark code, raw data and results: github.com/cezarc1/decodebench (see RESULTS.md).
MXFP4 W4A4 quantization of
Qwen/Qwen3-32B
(revision 9216db5781bf21249d130ec9da846c4624c16137, BF16). It was made for a controlled NVFP4-vs-MXFP4 decode
benchmark on NVIDIA B200 (fp4bench), not as a general-purpose release: the NVFP4 and MXFP4
checkpoints share the model, the tool, the recipe, the calibration data and the layer coverage,
and differ only in the format.
| Weights | MXFP4: FP4 E2M1 values in 32-element blocks, E8M0 (power-of-two) block scales, 4.25 bits per weight |
| Activations | 4-bit, quantized dynamically per 32-element block at runtime; this format has no static activation scales, so none are stored |
| Coverage | all Linear layers except lm_head (448 modules); embeddings, norms and lm_head stay BF16 |
| KV cache | BF16 at serve time (--kv-cache-dtype bfloat16); the checkpoint has no kv_cache_scheme |
How it was made
Configuration
- Architecture
- Qwen3ForCausalLM
- Context length (tokens)
- 40,960
- Layers
- 64
- Hidden size
- 5,120
- Feed-forward size
- 25,600
- Attention heads
- 64
- Key/value heads
- 8
- Head dimension
- 128
- Vocabulary size
- 151,936
- RoPE base
- 1,000,000
- Stored precision
- bfloat16
- Model type
- qwen3
- Quantization
- compressed-tensors
Identity and Version
- Repository
- ggamecrazy/Qwen3-32B-MXFP4-W4A4
- Publisher
- Cezar
- Task
- Text generation
- Modality
- Text
- Library
- Not stated by the source
- Parameters
- 32.8B parameters
- Languages
- Not stated by the source
- Revision
- a49f042511ac83042bfd39e8d5156eed98acca45
- First published
- 2026-10-05
- Last updated
- 2026-10-09
Files and Weights
12 files, 19.7 GB in total. The weights are 1 file totalling 19.7 GB in safetensors.
Weights1 file · 19.7 GB
Configuration4 files · 21.0 KB
Tokenizer4 files · 15.9 MB
Documentation2 files · 14.4 KB
Repository1 file · 1.6 KB
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| model.safetensors | Weights | 19.7 GB | 108d233a6354 |
| config.json | Configuration | 2.1 KB | — |
| fp4bench_quant.json | Configuration | 18.5 KB | — |
| generation_config.json | Configuration | 239 B | — |
| recipe.yaml | Configuration | 209 B | — |
| LICENSE | Documentation | 11.3 KB | — |
| README.md | Documentation | 3.0 KB | — |
| .gitattributes | Repository | 1.6 KB | — |
| merges.txt | Tokenizer | 1.7 MB | — |
| tokenizer.json | Tokenizer | 11.4 MB | aeb13307a71a |
| tokenizer_config.json | Tokenizer | 9.7 KB | — |
| vocab.json | Tokenizer | 2.8 MB | — |
License and Download
- License
- apache-2.0
- Access
- Open weights, no gate
- Download size
- 19.7 GB
Download from Cezar
Released by Cezar through its official repository on Hugging Face. Read the license.
Built From
- Derived from Qwen/Qwen3-32B
- Quantized from Qwen/Qwen3-32B
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 19.7 GB |
| 16-bit | 65.5 GB |
| 8-bit | 32.8 GB |
| 4-bit | 16.4 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Questions About Qwen3-32B-MXFP4-W4A4
How much GPU memory does Qwen3-32B-MXFP4-W4A4 need?
About 78.6 GB at 16-bit and 19.7 GB at 4-bit: the weights (32.8B parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run Qwen3-32B-MXFP4-W4A4 on?
At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.
Can I use Qwen3-32B-MXFP4-W4A4 commercially?
Yes. Qwen3-32B-MXFP4-W4A4 is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.
What is Qwen3-32B-MXFP4-W4A4's context length?
40,960 tokens, from the maximum position embeddings in its published configuration.
Similar Models
Qwen3 is the latest generation of large language models in Qwen series, offering a comprehensive suite of dense and mixture-of-experts (MoE) models. Built upon extensive training, Qwen3 delivers groundbreaking advancements in reasoning, instruction-following, agent capabilities, and multilingual support, with the following key features: - Uniquely support of seamless switching between thinking mode (for complex logical reasoning, math, and coding) and non-thinking mode (for efficient, general-purpose dialogue) within single model, ensuring optimal performance across various scenarios. - Significantly enhancement in its reasoning capabilities, surpassing previous QwQ (in thinking mode) and…
This qwen3 model was trained 2x faster with Unsloth and Huggingface's TRL library.
This qwen3 model was trained 2x faster with Unsloth and Huggingface's TRL library.
This qwen3 model was trained 2x faster with Unsloth and Huggingface's TRL library.
This qwen3 model was trained 2x faster with Unsloth and Huggingface's TRL library.