GLM-5.2-224E-FP8-REAP

GLM-5.2-224E-FP8-REAP

This is an independently produced, expert-pruned derivative of zai-org/GLM-5.2-FP8. It is not an official Z.ai release.

The checkpoint reduces each sparse MoE layer from 256 routed experts to 224 while preserving 8 active experts per token. Experts were selected with REAP scores computed from actual router weights and unweighted expert-output norms. The weights retain the source model's block-scaled FP8 format.

Model Details

Pruning Method

REAP calibration used 12,228 samples and 15,425,969 tokens across general, code, reasoning, and agentic data. For every routed expert, the score was the conditional mean of:

actual_router_weight * L2(unweighted_expert_output)

The 32 lowest-scoring experts in each of the 75 sparse layers were removed. The derivative was built directly from the immutable 256-expert source; it was not produced by repeatedly pruning another derivative.

Keep-map SHA-256: 6845c6d68d39934a5866b12b1222cd29da214b1080485cf39d8a9986b7c203f6

Validation

• Deep checkpoint validation: 104,037 of 104,037 tensors passed, with zero errors.
• Runtime functional checks: 4 of 4 passed, covering arithmetic, exact instruction following, long-context retrieval, and structured tool calling.
• Maximum context validated on the tested runtime: 131,072 tokens with BF16 KV cache.
• Aggregate context matrix passed at 16 x 8K, 8 x 16K, 4 x 32K, 2 x 64K, and 1 x 126K.
• Sequential decode: 20.8689 output tokens/s on the tested 8-HPU setup.
• Concurrent decode: 263.5073 aggregate output tokens/s for 16 requests with 128 output tokens each.

Preliminary GPQA Check

A 10-question GPQA Diamond smoke sample scored 9/10 (90.0%) with no request errors or malformed answers. The published base-model score is 91.2%, but this small sample is not statistically sufficient for an exact quality comparison. It used the OpenAI simple-evals zero-shot chain-of-thought prompt with temperature=1.0 and top_p=0.95.

Deployment

The checkpoint was qualified with vLLM-Gaudi using tensor parallel 8, expert parallel 8, BF16 KV cache, and MTP disabled. A representative configuration is:

vllm serve destr8803/GLM-5.2-224E-FP8-REAP \
  --tensor-parallel-size 8 \
  --enable-expert-parallel \
  --kv-cache-dtype auto \
  --max-model-len 131072

Hardware capacity, graph configuration, and supported context length depend on the inference backend. The architecture advertises a larger nominal context, but this derivative has only been runtime-qualified through 131,072 tokens.

GLM-5.2 uses thinking by default. Keep the default maximum reasoning effort for benchmark reproduction; explicitly configure the serving framework if a lower reasoning effort or disabled thinking is desired.

Limitations

• This is a structural MoE pruning derivative. Removing experts can affect quality unevenly across domains even when aggregate smoke tests pass.
• The full official GLM-5.2 benchmark suite has not been reproduced.
• The preliminary 10-question GPQA result must not be interpreted as a complete benchmark result.
• Only the BF16 KV-cache runtime profile is quality-qualified. FP8 KV cache was not accepted for this release.
• Deployment requires substantial storage and accelerator memory.

The base model's intended-use, safety, and language limitations continue to apply. Review the zai-org/GLM-5.2 model card before deployment.

License and Attribution

This derivative is distributed under the MIT license included in the repository. GLM-5.2 was created by Z.ai and Tsinghua University. Please cite the original technical report:

@misc{glm5team2026glm5vibecodingagentic,
  title={GLM-5: from Vibe Coding to Agentic Engineering},
  author={GLM-5-Team},
  year={2026},
  eprint={2602.15763},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2602.15763}
}