GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
zhudanburujiandan
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
zhudanburujiandan
Model Tree
Qwen/Qwen2-VL-7B-Instruct
Input Modalities
Output Modalities
Supported Functionality
| Metric (100 val QAs, seed=42) | v1 (154k, no speed) | v3-20k (this, speed) | v3 (154k, speed) |
|---|---|---|---|
| Exact | 62 | 64 | 66 |
| Perception | 16 | 13 | 16 |
| Prediction | 22 | 22 | 22 |
| Planning | 13 | 12 | 13 |
| Behavior | 11 (44%) | 17 (68%) ⭐ | 15 (60%) |
Pick this variant when:
Pick v3-full when:
Identical to v3 — see the v3 README for the full example. Just swap the adapter path to this repo.
seed=42).lr=2e-5, cosine schedule, bsz=1, grad-accum=8, bf16.eval_loss plateau 0.184 (v3-full 0.175, essentially matched).Speculation, not proven: with less data the model over-relies on the numeric speed prefix (which perfectly predicts the "fast/slow/…" bin) and under-fits the tiny visual variations that push it to output non-canonical answers. At 154k the model sees enough perception / prediction / planning examples to develop a more "distributed" style that occasionally deviates from the terse behavior template.
Practical takeaway: if you're deploying, keep both variants around and pick per-QA-level, or use v3-full which is the safer overall bet. This variant is included for reproducibility of the finding.
Everything else (base model, dataset, license, limitations, citation) — see v3.