GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
intel-ai
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
intel-ai
Model Tree
zai-org/GLM-5.3
Input Modalities
Output Modalities
Supported Functionality
This model is an MXFP4 (W4A4, group size 32) build of
zai-org/GLM-5.3 produced with
intel/auto-round (--model_free RTN), with module
coverage identical to RedHatAI/GLM-5.3-MXFP4.
Please follow the license of the original model.
q_a_proj, q_b_proj, kv_a_proj_with_mqa, kv_b_proj, o_proj), shared expert
projections, the leading dense MLPs (layers 0-2) and the DSA indexer's wq_b.wk and weights_proj, MoE router (mlp.gate), the MTP
fusion projection (eh_proj), token embedding, LM head, LayerNorms.total_size 403,321,862,144).
The RedHatAI repo's whole-repo figure is 403,368,910,311 bytes (403.37 GB) — the extra 31.8 MB
is config/tokenizer/README, not weights.bash
# produced and validated with:# vllm 0.32.1rc1.dev253+g30aded62d (registers GlmMoeDsaForCausalLM via vllm.models.deepseek_v32)# torch 2.14.0+cu130 | auto-round 0.16.1 | compressed-tensors 0.17.0 | lm-eval 0.4.13uv pip install "vllm==0.32.1rc1.dev253"bash
vllm serve {org}/GLM-5.3-MXFP4-CT-RTN-AutoRound \ --tensor-parallel-size 4 \ --dtype bfloat16 \ --reasoning-parser glm45--tensor-parallel-size 4 is what this artifact was validated at (403 GB → ~101 GB/GPU).
Other topologies are fine as long as moe_intermediate_size / (TP * 32) stays a multiple of 4
(2048 → TP=1/2/4/8 all satisfy it).
lm_eval 0.4.13 + vLLM 0.32.1rc1.dev253, TP=4, enable_thinking=False, seed 42.
PIQA / HellaSwag / MMLU loglikelihood, GSM8K chat template with 5-shot; full test sets, no --limit.
| Benchmark | This model | |---|---:|---:| | PIQA (acc) | 83.46 | | HellaSwag (acc) | 71.50 | | MMLU (acc) | 84.23 | | GSM8K (strict-match) | 97.04 | | Average | 84.06 |
Two ingredients: the auto-round CLI invocation below, and a small opt-in patch that disables
auto-round's built-in GLM ignore rule (auto_round/special_model_handler.py) so the dense MLPs
and indexer.wq_b are quantized like RedHatAI does.
bash
# 1) patch auto-round in a private overlay (shared venv is left untouched)SITE=<venv>/lib/python3.12/site-packagesmkdir -p /path/to/ar_overlay && cp -r "$SITE/auto_round" /path/to/ar_overlay/patch --batch --fuzz=0 -p1 -d /path/to/ar_overlay -i patches/autoround_glm_moe_dsa_rh_strict.patch # 2) quantizePYTHONPATH=/path/to/ar_overlay AUTO_ROUND_GLM_RH_STRICT=1 auto-round \ --model_name zai-org/GLM-5.3 \ --model_free \ --scheme MXFP4 \ --ignore_layers "mlp.gate$,lm_head,embed_tokens,eh_proj,indexer.weights_proj,indexer.wk" \ --format llm_compressor \ --device_map auto \ --output_dir <output_dir>mxfp4-pack-quantized — weights E2M1, group size 32, E8M0 scales;
activations FP4 with dynamic per-group scaling (W4A4).["Linear"] (same as RedHatAI).mlp.gate$ (MoE router; the $ anchor matters
because the dense layers have mlp.gate_proj), lm_head, embed_tokens, eh_proj,
indexer.weights_proj, indexer.wk. Expands to 123 module names.text
ignore_layers=[get_glm_flash_ignore_layers, "weights_proj", "indexer"]glm_moe_dsa family, which keeps the leading dense MLPs
and the entire DSA indexer in BF16. Those workarounds target the tuning/forward path
(torch.compile AOTAutograd with the indexer's torch.no_grad(), and a vLLM loading issue);
with --model_free nothing is loaded or compiled, so the narrower ignore is safe.
Without the patch the artifact is 31 tensors larger in BF16 terms and ~1.3 GB bigger.scripts/glm53_quant_rh_style.sh (aqa quant.command_override).The model can produce factually incorrect output, and should not be relied on to provide factually accurate information. Because of the limitations of the pretrained model and the finetuning datasets, it is possible that this model could generate lewd, biased or otherwise offensive outputs. Therefore, before deploying any applications of this model, developers should perform safety testing.
Users (both direct and downstream) should be made aware that the risks of this model involve hallucinations, biased or offensive content, and factual errors.
bibtex
@article{cheng2023optimize, title={Optimize weight rounding via signed gradient descent for the quantization of large language models}, author={Cheng, Wenhua and Zhang, Weiwei and Shen, Haihao and Cai, Yiyang}, journal={arXiv preprint arXiv:2309.05516}, year={2023}}The license on this model does not constitute legal advice. We are not responsible for the actions of third parties using this model. Please consult an attorney if you need legal advice about your use of this model, and follow the license of zai-org/GLM-5.3.
Signed-off-by: Kaokao, Lv kaokao.lv@intel.com
# produced and validated with:# vllm 0.32.1rc1.dev253+g30aded62d (registers GlmMoeDsaForCausalLM via vllm.models.deepseek_v32)# torch 2.14.0+cu130 | auto-round 0.16.1 | compressed-tensors 0.17.0 | lm-eval 0.4.13uv pip install "vllm==0.32.1rc1.dev253"vllm serve {org}/GLM-5.3-MXFP4-CT-RTN-AutoRound \ --tensor-parallel-size 4 \ --dtype bfloat16 \ --reasoning-parser glm45# 1) patch auto-round in a private overlay (shared venv is left untouched)SITE=<venv>/lib/python3.12/site-packagesmkdir -p /path/to/ar_overlay && cp -r "$SITE/auto_round" /path/to/ar_overlay/patch --batch --fuzz=0 -p1 -d /path/to/ar_overlay -i patches/autoround_glm_moe_dsa_rh_strict.patch # 2) quantizePYTHONPATH=/path/to/ar_overlay AUTO_ROUND_GLM_RH_STRICT=1 auto-round \ --model_name zai-org/GLM-5.3 \ --model_free \ --scheme MXFP4 \ --ignore_layers "mlp.gate$,lm_head,embed_tokens,eh_proj,indexer.weights_proj,indexer.wk" \ --format llm_compressor \ --device_map auto \ --output_dir <output_dir>@article{cheng2023optimize, title={Optimize weight rounding via signed gradient descent for the quantization of large language models}, author={Cheng, Wenhua and Zhang, Weiwei and Shen, Haihao and Cai, Yiyang}, journal={arXiv preprint arXiv:2309.05516}, year={2023}}