GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
HenryExcellent
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
HenryExcellent
Model Tree
Qwen/Qwen3.6-27B
Input Modalities
Output Modalities
Supported Functionality
Merged model weights accompanying SciDocBench. GRPO checkpoint at optimizer step 200, initialized from the merged SFT step-50 model.
The backbone is Qwen/Qwen3.6-27B. These are complete merged weights, not a LoRA-only adapter. Use the model's tokenizer, chat template, and image processor with a compatible Transformers or vLLM installation.
During evaluation, retain the generated response losslessly and score the final answer after reasoning extraction. See the paper for the experiment-specific training mixtures, inference settings, and checkpoint selection protocol.
upload_manifest.json and SHA256SUMS identify the files in this model release.
The base model's Apache-2.0 license is included. Local training logs and private
runtime configuration are not part of this repository.