Modelpublic

mocoV3/R2EAct-Qwen3.8-27B-Compressor-r16-step1000

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes24downloads

Model Card

R2EAct Qwen3.8-27B Compressor (r16, step 1000)

This repository contains the deployment checkpoint for a recursive context compressor trained against Qwen3.8-27B.

It is a custom compressor checkpoint, not a standalone Transformers language model. The frozen Qwen3.8-27B target weights are not included; use `Qwen/Qwen3.8-27B` separately with the compatible R2EAct/SpecForge SGLang integration.

Contract

—Target family: Qwen3.8-27B multimodal
—Compression ratio: 1:16
—Training checkpoint: step 1000
—Serving schema: qwen38_27b_r2eact_token_clock_serving_v1
—Maximum encoder sequence: 12,288 tokens
—Compressor SHA-256: 6137e1da636e6d1b1805d3018bcc4f7388eab25469569d6eca4007c91019b3c1
—Canonical config SHA-256: 54484fda900036b49a8b03dc385e022e5d6b61f27c62768ff48f35f168ed7f86

Files

—compressor.pt: compressor state dictionary
—config.json: architecture, training provenance, and serving contract
—manifest.json: integrity metadata

The paths retained in config.json are provenance from the training/export environment. Deployment should explicitly bind the target model to Qwen/Qwen3.8-27B or a verified local snapshot.