Original title: KORMo-VL/KORMo-Web-11B
Uploaded by the KAIST MLP Lab on 11/14, based on their KORMo-10B model, which was pre-trained from scratch for Korean and English. It was extended with KORMo-VL (SigLIP vision) and then retrained as a web agent: the same model can converse, view images, and operate a browser using only screenshots (22 types of actions: click, input, keyboard shortcuts, scroll, drag, switch tabs, etc.), writing its reasoning before outputting each action. The authors state they used over 1,0000,000 web understanding samples and over 100,000 Korean-English operation trajectories. Licensed under Apache-2.0; the KORMo-Web-9B model in the same series, based on Qwen, specializes in web tasks.
Why it matters It operates browsers while retaining chat and image-viewing capabilities; the authors claim it has the highest success rate for Korean web tasks among open models listed in the table
Upvotes 70 · Likes 77 · Downloads 26
Reasons
- The authors evaluated using the same model on 612 real website tasks (English 300, Korean 312) in Online-Mind2Web: Korean success rate 30.1%, English 29.0%; in the same table, UI-TARS-1.5-7B has a Korean success rate of 20.2%, and Qwen3.5-9B has 16.7%
- Uses only screenshots, requiring no DOM or accessibility tree; the repository includes full message format documentation (HARNESS.md), 22 tool definitions, and a Playwright reference agent
- Can run on a single GPU; loading paths tested by the authors are provided for both transformers 5.12.1 and vLLM 0.23.0
Usage
- transformers: After downloading the repository, first import kormo_load and call kormo_load.register(path), then load AutoProcessor and AutoModelForImageTextToText. After loading, call kormo_load.patch_vision(model) (failing to do so will result in incorrect screenshot reading)
- vLLM: Convert the format once using vllm/convert_to_vllm.py, set PYTHONPATH to the repository's vllm directory, then run vllm serve; use examples/agent_playwright.py --task "..." to have it open Chromium and perform tasks
- Recommended temperature 0.6, top_p 0.95, top_k 20, with thinking enabled, maximum 16k tokens per step, and maximum 50 steps (already set in generation_config); output coordinates are normalized values between 0 and 1000
Risks
- Web task success rate is around 30%, a large gap compared to closed-source models in the table (over 60%); the authors state that specialized versions of the same series, 4B / 9B, have higher web scores
- Supports only Korean and English; success rates are automatically judged by GPT-6, with no manual verification
- Having the model directly operate a browser may lead to misclicks or erroneous submissions; it is recommended to run in an isolated browser environment. The reference agent does not execute uploads. This entry has not been run