Akicou/GLM-4.7-Flash-REAP-39

𓌳 REAP𓌳 the Experts: Why Pruning Prevails for One-Shot MoE Compression
📄 Paper • 💻 Code • 📝 Blog

GLM-4.7-REAP-39

⚠️ Model Status & Deployment Note

Important Update regarding GGUF support: A critical bug was recently identified in llama.cpp regarding the scoring_func for GLM models (which caused looping and poor output quality). While the base weights are functional, the GGUF files for this model are currently being re-generated to ensure full compatibility with the latest fixes.

• Status: GGUF files are scheduled for re-upload by January 24, 2026.
• Recommendation: If you are using local inference via llama.cpp or Unsloth, please refer to the official Unsloth GLM-4.7-Flash documentation for the most stable configuration parameters.
• Native Support: The BF16/FP16 weights remain compatible with transformers and vLLM for immediate use.

✨ Highlights

50% Expert-Pruned GLM-4.7 Flash optimized for code generation, function calling, and agentic workflows.

Created using REAP (Router-weighted Expert Activation Pruning) by Cerebras:

• Calibrated for Code & Tools: Preserves coding and function-calling capabilities
• One-Shot Compression: No fine-tuning required
• Drop-in Compatible: Works with vLLM, Transformers, SGLang

🙏 Acknowledgments

• Runpod — Compute for REAP
• Cerebras — REAP methodology

--

The Science Behind Dataset Selection

REAP Algorithm:
1. Forward pass calibration samples through model
2. Record which experts activate and their magnitudes
3. Compute saliency = router_weight × activation_norm
4. Prune lowest-saliency experts

Key Insight: Experts are TASK-SPECIFIC
├── Some experts specialize in natural language
├── Some experts specialize in code syntax
├── Some experts specialize in JSON/structured output
└── Some experts specialize in multi-turn context

If calibration lacks code → code-specialized experts appear "unused" → get pruned → model loses coding ability

Cerebras' Original Mix (from paper)

Cerebras used the same 3 datasets in their GLM-4.6 REAP experiments:

• evol-codealpaca-v1 for code generation
• xlam-function-calling-60k for tool calling
• SWE-smith-trajectories for agentic tasks

We followed this exact recipe for reproducibility.

---

🚀 Deployment

vLLM (Recommended)

vllm serve Akicou/GLM-4.7-Flash-REAP-39 \
    --tensor-parallel-size 2 \
    --trust-remote-code \
    --dtype bfloat16

Transformers

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "Akicou/GLM-4.7-Flash-REAP-39",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("Akicou/GLM-4.7-REAP-39", trust_remote_code=True)

messages = [{"role": "user", "content": "Write a Python function to merge two sorted lists."}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True)
outputs = model.generate(inputs.to(model.device), max_new_tokens=512, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

⚖️ License

MIT (inherited from GLM-4.7 Flash)

---

🧾 Citation

@article{lasby2025reap,
  title={REAP the Experts: Why Pruning Prevails for One-Shot MoE Compression},
  author={Lasby, Mike and Lazarevich, Ivan and Sinnadurai, Nish and Lie, Sean and Ioannou, Yani and Thangarasa, Vithursan},
  journal={arXiv preprint arXiv:2510.13999},
  year={2025},
  url={https://arxiv.org/abs/2510.13999}
}