GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
mustafaah
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
mustafaah
Model Tree
Qwen/Qwen3-VL-2B-Instruct
Input Modalities
Output Modalities
Supported Functionality
Qwen3-VL-2B-Instruct with our highlight-only SFT + RL update-200 LoRA merged in BF16. This is the trained model, not the base checkpoint. It returns one JSON highlight response for a screenshot and instruction.
Base revision: 89644892e4d85e24eaac8bacfd4f463576704203.
Adapter SHA256: 10c98d1751b609c2c0a7c72cf404f07b931f2aea7220e32106078c0a4390c9ee.
Adapter source: final RL adapters.
The included merge validation compares the same screenshot before/after merging: six corresponding navigation boxes, at most one normalized-coordinate unit difference. BF16 merging is not token-exact. This smoke test is not a full evaluation.
Use Transformers 4.57.6 and the system prompt included here. Input: full screenshot plus instruction; output: {"op":"highlight","targets":[["yellow",x0,y0,x1,y1],...]} in 0–1000 coordinates. Greedy decoding; max 1024 output tokens; 256–1024 image tokens. No external OCR or inspection tools.