Laguna-S-2.1-oQ2e-fast-REAP-164

Laguna S 2.1 oQ2e-fast REAP-164

This is an MLX checkpoint derived from poolside/Laguna-S-2.1. It applies REAP expert pruning to the oQ2e-fast quantized model while retaining the original model's architecture, tokenizer, chat template, and mixed-precision quantization layout.

Model details

• Architecture: Laguna mixture-of-experts causal language model
• Sparse transformer layers: 47
• Routed experts retained per sparse layer: 164 of 256
• Active routed experts per token: 10
• Dense layer 0 is preserved
• Shared experts are preserved in every sparse layer
• Routed scaling factor: 2.5
• Format: MLX SafeTensors

REAP ranks routed experts from calibration activations and removes the lowest-saliency experts independently in each sparse layer. The retained experts keep their original ordering and routing behavior.

Usage

The checkpoint includes the Laguna model implementation, tokenizer, generation configuration, and chat template required by MLX-compatible runtimes. Load the repository as an MLX model and apply the bundled chat template for conversational generation and tool-calling prompts.

Base model and terms

This derivative follows the license and usage terms of poolside/Laguna-S-2.1. Review the base model card for its intended uses, limitations, and license conditions.

Acknowledgements

• Base model: poolside/Laguna-S-2.1
• REAP: CerebrasResearch/reap