GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
ohjoonhee
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
ohjoonhee
Model Tree
Qwen/Qwen2.5-VL-7B-Instruct
Input Modalities
Output Modalities
Supported Functionality
Qwen/Qwen2.5-VL-7B-Instruct(none)ohjoonhee/visual-cot-50k-poc (Monet-SFT-125K Visual_CoT subset, eval-200 excluded)Pure NTP SFT — no Monet Stage 2 alignment loss, no latent-mode forward.
The Monet special tokens (<observation>, <abs_vis_token>, etc.) ARE
registered in the tokenizer and embedded so the model learns to produce
them, but the architectural latent-slot mechanism is unused at this stage.
step-2913)No training log row available.
Faithful epochs=3 upstream Monet Stage 3 (released recipe). 2-job resume (218542+218752), full 3 epochs / step 2913. Init Monet-SFT-7B/stage1, teacher upstream-precomputed 124K latents. Final train align_loss 0.026, obs_acc 0.99.
Other revisions: see the revisions dropdown on this page.
python
from transformers import AutoModelForVision2Seq, AutoProcessorm = AutoModelForVision2Seq.from_pretrained("ohjoonhee/vlatents-qwen25vl7b-stage3-upstream-ep3-v1", revision="step-2913", torch_dtype="bfloat16")p = AutoProcessor.from_pretrained("ohjoonhee/vlatents-qwen25vl7b-stage3-upstream-ep3-v1", revision="step-2913")
Research checkpoint, eval-only. Mid-training step (2913/?). Not for production.
Card generated 2026-06-04 from training_log.jsonl + the run's training config.