GLM-4.6-REAP-268B-A32B-Derestricted-FP8

<div align="left"> <img src=https://cdn-uploads.huggingface.co/production/uploads/6625f4a8a8d1362ebcc3851a/iyzgR89q50pp1T8HeeP15.png width="5%"/> </div>

Arli AI

GLM-4.6-REAP-268B-A32B-Derestricted

<div align="center"> <img src=https://cdn-uploads.huggingface.co/production/uploads/6625f4a8a8d1362ebcc3851a/XhCz9N4liIwWEh-yH37gR.png width="15%"/> </div>

GLM-4.6-REAP-268B-A32B-Derestricted is a Derestricted version of GLM-4.6-REAP-268B-A32B, created by Arli AI.

Our goal with this release is to provide a version of the model that removed refusal behaviors while maintaining the high-performance reasoning of the original GLM-4.5-Air. This is unlike regular abliteration which often inadvertently "lobotomizes" the model.

Methodology: Norm-Preserving Biprojected Abliteration

To achieve this, Arli AI utilized Norm-Preserving Biprojected Abliteration, a refined technique pioneered by Jim Lai (grimjim). You can read the full technical breakdown in this article.

Why this matters:

Standard abliteration works by simply subtracting a "refusal vector" from the model's weights. While this works to uncensor a model, it is mathematically unprincipled. It alters the magnitude (or "loudness") of the neurons, destroying the delicate feature norms the model learned during training. This damage is why many uncensored models suffer from degraded logic or hallucinations.

How Norm-Preserving Biprojected Abliteration fixes it:

This model was modified using a three-step approach that removes refusals without breaking the model's brain:

1. Biprojection (Targeting): We refined the refusal direction to ensure it is mathematically orthogonal to "harmless" directions. This ensures that when we cut out the refusal behavior, we do not accidentally cut out healthy, harmless concepts.
2. Decomposition: Instead of a raw subtraction, we decomposed the model weights into Magnitude and Direction.
3. Norm-Preservation: We removed the refusal component solely from the directional aspect of the weights, then recombined them with their original magnitudes.

The Result:

By preserving the weight norms, we maintain the "importance" structure of the neural network. Benchmarks suggest that this method avoids the "Safety Tax"—not only effectively removing refusals but potentially improving reasoning capabilities over the baseline, as the model is no longer wasting compute resources on suppressing its own outputs.

In fact, you may find surprising new knowledge and capabilities that the original model does not initially expose.

Quantization:

• Original: https://huggingface.co/ArliAI/GLM-4.6-REAP-268B-A32B-Derestricted
• FP8: https://huggingface.co/ArliAI/GLM-4.6-REAP-268B-A32B-Derestricted-FP8
• INT8: https://huggingface.co/ArliAI/GLM-4.6-REAP-268B-A32B-Derestricted-W8A8-INT8

---

Original model card:

---

language:

• en library_name: transformers tags:
• glm
• MOE
• pruning
• compression license: mit name: cerebras/GLM-4.6-REAP-268B-A32B description: > This model was obtained by uniformly pruning 25% of experts in GLM-4.6 using the REAP method. readme: > https://huggingface.co/cerebras/GLM-4.6-REAP-268B-A32B/main/README.md license_link: https://huggingface.co/zai-org/GLM-4.6/blob/main/LICENSE pipeline_tag: text-generation base_model:
• zai-org/GLM-4.6

---

<p align="center"> <em>𓌳 <strong>REAP</strong>𓌳 the Experts: Why Pruning Prevails for One-Shot MoE Compression</em><br> <img src="https://i.imgur.com/rmzG3gg.png" alt="REAP" width="75%"> </p>

GLM-4.6-REAP-268B-A32B

✨ Highlights

Introducing GLM-4.6-REAP-268B-A32B, a memory-efficient compressed variant of GLM-4.6 that maintains near-identical performance while being 25% lighter.

Note: this is a BF16 version for more accurate downstream low-bit quantization. An FP8 version is also available on HF.

This model was created using REAP (Router-weighted Expert Activation Pruning), a novel expert pruning method that selectively removes redundant experts while preserving the router's independent control over remaining experts. Key features include:

• Near-Lossless Performance: Maintains almost identical accuracy on code generation, agentic coding, and function calling tasks compared to the full 355B model
• 25% Memory Reduction: Compressed from 355B to 268B parameters, significantly lowering deployment costs and memory requirements
• Preserved Capabilities: Retains all core functionalities including code generation, agentic workflows, repository-scale understanding, and function calling
• Drop-in Compatibility: Works with vanilla vLLM - no source modifications or custom patches required
• Optimized for Real-World Use: Particularly effective for resource-constrained environments, local deployments, and academic research

---

📋 Model Overview

GLM-4.6-REAP-268B-A32B has the following specifications:

• Base Model: GLM-4.6
• Compression Method: REAP (Router-weighted Expert Activation Pruning)
• Compression Ratio: 25% expert pruning
• Type: Sparse Mixture-of-Experts (SMoE) Causal Language Model
• Number of Parameters: 268B total, 32B activated per token
• Number of Layers: 92
• Number of Attention Heads (GQA): 96 for Q and 8 for KV
• Number of Experts: 120 (uniformly pruned from 160)
• Number of Activated Experts: 8 per token
• Context Length: 202,752 tokens
• License: MIT

---

📊 Evaluations

TBD for BF16 model. Evalulation results available for the FP8 variant.

For more details on the evaluation setup, refer to the REAP arXiv preprint.

---

🚀 Deployment

You can deploy the model directly using the latest vLLM (v0.11.0), no source modifications or custom patches required.

vllm serve cerebras/GLM-4.6-REAP-268B-A32B \
    --tensor-parallel-size 8 \
    --tool-call-parser glm45 \
    --enable-auto-tool-choice \
    --enable-expert-parallel

If you encounter insufficient memory when running this model, you might need to set a lower value for --max-num-seqs flag (e.g. set to 64).

🧩 Model Creation

This checkpoint was created by applying the REAP (Router-weighted Expert Activation Pruning) method uniformly across all Mixture-of-Experts (MoE) blocks of GLM-4.6, with a 25% pruning rate.

How REAP Works

REAP selects experts to prune based on a novel saliency criterion that considers both:

• Router gate values: How frequently and strongly the router activates each expert
• Expert activation norms: The magnitude of each expert's output contributions

This dual consideration ensures that experts contributing minimally to the layer's output are pruned, while preserving those that play critical roles in the model's computations.

Key Advantages

• One-Shot Compression: No fine-tuning required after pruning - the model is immediately ready for deployment
• Preserved Router Control: Unlike expert merging methods, REAP maintains the router's independent, input-dependent control over remaining experts, avoiding "functional subspace collapse"
• Generative Task Superiority: REAP significantly outperforms expert merging approaches on generative benchmarks (code generation, creative writing, mathematical reasoning) while maintaining competitive performance on discriminative tasks

Calibration

The model was calibrated using a diverse mixture of domain-specific datasets including:

• Code generation samples (evol-codealpaca)
• Function calling examples (xlam-function-calling)
• Agentic multi-turn trajectories (SWE-smith-trajectories)

📚 For more details, refer to the following resources:

• 🧾 arXiv Preprint
• 🧾 REAP Blog
• 💻 REAP Codebase (GitHub)

---

⚖️ License

This model is derived from zai-org/GLM-4.6 and distributed under the MIT license.

---

🧾 Citation

If you use this checkpoint, please cite the REAP paper:

@article{lasby-reap,
  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}
}