GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
sjin59
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
sjin59
Model Tree
Qwen/Qwen3.5-27B
Input Modalities
Output Modalities
Supported Functionality
Intermediate checkpoint, not the final model. This is the SFT-stage checkpoint that the
GRPO stage started from, released so the training pipeline can be resumed from the middle.
For inference use sjin59/c43g-s600m2.
The task: given 4 shuffled frames from a video plus a caption describing what happens,
recover the original chronological order. Output is four letters, e.g. "C,B,A,D".
Full supervised fine-tune of Qwen3.5-27B with the vision tower and aligner frozen — 16,070 rows, 2 epochs (1,004 steps), seed 43, effective batch 32, lr 3e-6 cosine, ZeRO-3.
Runs on vLLM 0.19.0. Licensed under Apache-2.0, following the base model Qwen3.5-27B.