Missing MXFP4 quantized models for Qwen3-235B-A22B, MiniMax-M2.5, and GLM-5

[sunway513]

Summary

Several high-priority frontier MoE models are missing from the AMD Quark MXFP4 quantized models collection on HuggingFace (amd/ org). NVIDIA has already published NVFP4 versions for all of these, creating a benchmark gap.

Missing Models

Model	HuggingFace ID	Params	Active	NVIDIA NVFP4 Available?
Qwen3-235B-A22B	Qwen/Qwen3-235B-A22B	235B	22B	✅ nvidia/Qwen3-235B-A22B-NVFP4
MiniMax-M2.5	MiniMaxAI/MiniMax-M2.5	229B	10B	✅ nvidia/MiniMax-M2.5-NVFP4
GLM-5	zai-org/GLM-5	745B	44B	✅ nvidia/GLM-5-NVFP4

Existing AMD MXFP4 Models (for reference)

The following models already have MXFP4 versions:

• amd/DeepSeek-R1-0528-MXFP4 ✅
• amd/Kimi-K2.5-MXFP4 ✅
• amd/GLM-5-MXFP4 ✅ (exists but has weight shape issues in ATOM — see below)
• amd/Qwen3.5-397B-A17B-MXFP4 ✅
• amd/Qwen3-Coder-Next-MXFP4 ✅

Why This Matters

1. Benchmark parity: Without MXFP4 versions, MI355X benchmarks must use FP8 (2x larger, 2x less compute density), making MI355X vs B300 comparisons unfair
2. InferenceMAX/InferenceX: SemiAnalysis benchmarks use NVFP4 on NVIDIA side — AMD needs equivalent MXFP4 for apples-to-apples comparison
3. Customer readiness: These are top-5 deployed open MoE models as of March 2026

Additional Issue: GLM-5 MXFP4 Weight Shape Bug

amd/GLM-5-MXFP4 exists on HuggingFace but fails to load in ATOM with:

RuntimeError: The size of tensor a (3072) must match the size of tensor b (12288) at non-singleton dimension 1

This occurs in the weight loader during MXFP4 unpacking with TP=8. The GLM-5 hidden_size=6144 divided by TP=8 gives 768, but the MXFP4 packed tensor dimension is 12288 (expecting undivided size). Please verify the quantization output format.

Request

1. Publish amd/Qwen3-235B-A22B-MXFP4 (or Instruct variant)
2. Publish amd/MiniMax-M2.5-MXFP4
3. Fix and re-publish amd/GLM-5-MXFP4 (weight shape issue)

cc @quark-team

[sunway513]

@thiagocrepaldi — Could you take a look at this? We're running MI355X vs B300 inference benchmarks for frontier MoE models and the missing MXFP4 weights are blocking fair comparisons. NVIDIA already has NVFP4 versions for all of these on HuggingFace.

Priority order:

1. Qwen3-235B-A22B — highest enterprise demand, no MXFP4 safetensors exist anywhere
2. MiniMax-M2.5 — needed for competitive benchmarks
3. GLM-5 MXFP4 fix — amd/GLM-5-MXFP4 exists but has weight shape mismatch in ATOM

Any ETA or the right team to route this to?

[niuxjamd]

Hi, @sunway513,

These 3 models are available on Huggingface:
https://huggingface.co/amd/Qwen3-235B-A22B-Instruct-2507-MXFP4
https://huggingface.co/amd/MiniMax-M2.5-MXFP4
https://huggingface.co/amd/GLM-5-MXFP4

Will put resource in my team to validate these models in ATOM.