GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
isaacmg
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
isaacmg
Model Tree
Qwen/Qwen3-VL-8B-Instruct
Input Modalities
Output Modalities
Supported Functionality
| Model | What changed | Headline result |
|---|---|---|
| v1 | First fine-tune, printed Vilna Talmud | Reads unseen gemara at CER ~0.06–0.2 where the base model refused |
| v1.5 | + synthetic Rashi-script training | ~99% char accuracy on rendered Rashi (pure perception); crop-level gemara solved |
| v1.6 | Page-level training at 6.5 MP | Printed Rashi CER 0.047 — ahead of the frontier VLMs we benchmarked |
| v1.7 | + Cairo Genizah handwriting (curated corpus v1) | Best full-coverage Talmud model; first real handwriting capability |
| v1.8a | Repaired training data (ablation control) | Isolates the data-quality contribution |
| v1.8b | Vision tower actually trained (LoRA) | Outperforms the Kraken HTR baseline on our Genizah benchmark; no forgetting of printed scripts |
HebVL is a family of fine-tuned Qwen3-VL-8B adapters that read historical Hebrew-script material — from printed Talmud pages (square type, Rashi script) to the handwritten fragments of the Cairo Genizah: medieval letters, court deeds, and lists written in Hebrew, Judaeo-Arabic, and Aramaic, in the notoriously difficult Oriental documentary cursive of 11th–13th-century Egypt.
These models are part of the Cairo Genizah AI Project — a search engine, map, and semantic visualization of the Cairo Genizah corpus. Today, search there runs mostly on catalog metadata, because the great majority of Genizah fragments have never been transcribed. This model line exists to change that: once transcription quality is good enough, we will apply it across the untranscribed corpus to unlock genuine full-text and semantic search over one of the most important documentary archives of the medieval Mediterranean.
Where the line stands (v1.8b): on our frozen 131-fragment Genizah benchmark (verified image/transcription pairs, scored by behaviour classification, aligned F1, and order-independent n-gram precision), v1.8b is the first model in the line to outperform a specialist HTR baseline (Kraken with a medieval-Hebrew model): aligned F1 0.807 vs 0.730, with ink-fidelity (n-gram precision) at parity. On printed Talmud it transcribes gemara and Rashi script at character error rates competitive with or better than commercial frontier VLMs. An accompanying paper is under peer review; links will follow at camera-ready.
Honest limitations — please read before scholarly use. These are research
artifacts. On damaged or highly cursive fragments the models can produce fluent
but wrong text — invented dates, names, or filled-in lacunae that look like
readings. Never treat an output as an authoritative transcription without
checking it against the image. Practical tips: feed high-resolution scans (the
exported processors carry the training resolution policy, ~6.5 MP); expect better
results when prompting for a page or region at a time; damage gaps are marked
[...] by convention