model release

KAIST MLP Lab releases KORMo-Web-11B: open LLM, VLM and web agent scoring 30.1 on Korean Online-Mind2Web

TL;DR

KAIST MLP Lab has released KORMo-Web-11B, an 11B-parameter bilingual model that combines text, vision and browser-agent capabilities in one set of weights under Apache-2.0. On 612 Online-Mind2Web tasks it scores 29.0 in English and 30.1 in Korean, according to the lab.

3 min read
0

KAIST MLP Lab has released KORMo-Web-11B, an 11B-parameter Korean-English model that works as a language model, a vision-language model and a browser GUI agent. It scores 29.0 (English) and 30.1 (Korean) on Online-Mind2Web, according to the lab. Weights and supporting resources are published on Hugging Face under the Apache-2.0 license, which permits commercial use, modification and redistribution subject to the license terms. The model card is written in Korean and English.

Architecture and training

KORMo-Web-11B is built in three stages on the KORMo family, which KAIST MLP Lab trained from scratch:

  1. KORMo-10B: a Korean-English LLM with a Korean-optimized tokenizer.
  2. KORMo-VL: multimodal understanding through a SigLIP vision encoder and the KORMo language backbone.
  3. KORMo-Web-11B: web pretraining and action learning for visual browser interaction.

The lab says the web stage used 10M+ Korean and English pretraining samples covering UI elements, their locations, layouts, functions and states. It also used 100K+ bilingual web interaction trajectories for multi-step browser tasks.

The model operates primarily from screenshots and does not require full DOM or accessibility trees. It generates an inspectable reasoning step before each browser action.

The lab distinguishes this model from its Qwen-based KORMo-Web-4B and 9B models, which focus on maximizing web task performance. KORMo-Web-11B is designed to keep its original LLM and VLM capabilities while adding GUI control. The lab acknowledges that the 4B and 9B models achieve higher web task success.

Benchmark results

The lab evaluated on 612 live-website tasks from Online-Mind2Web (300 English, 312 Korean), using the same LLM judge for all models. Each agent uses its own prompt and action format. Success rates are computed over the full task set, including failed or incomplete runs.

Model Size EN KO
Qwen3.5-9B (base) 9B 24.7 16.7
Fara-1.5-4B 4B 20.7 7.7
UI-TARS-1.5-7B 7B 22.3 20.2
OpenWebRL-4B 4B 32.7 15.4
KORMo-Web-11B 11B 29.0 30.1
GPT-6-luna (proprietary) not listed 63.0 65.7

The lab states that the 30.1% Korean success rate exceeds the strongest non-KORMo open-weight baseline in the table by 9.9 percentage points. In English, OpenWebRL-4B scores higher (32.7 versus 29.0). The proprietary GPT-6-luna scores higher than every open model listed in both languages.

These are the lab's own evaluations, not independent results.

Running it

The model runs on a single GPU. The lab lists testing with transformers 5.12.1 and vLLM 0.23.0, and the repository includes a browser harness, reference inference settings, a vLLM inference path and Transformers 5 support. Transformers does not ship the KORMo text model, so the repository includes it (_modeling_kormo.py) with a loader, kormo_load.py. The lab says that without the patch_vision call the model still runs but misreads screenshots. vLLM needs a one-time conversion for the KORMo architecture.

The lab also provides a KORMo-Web Chrome extension that takes requests in Chrome's side panel and can be controlled from Discord. For Apple silicon, the documented path serves a 4-bit MLX build of KORMo-Web-4B with vllm-mlx, which the lab describes as simple but slower. On NVIDIA GPUs the extension points to the 11B model served through vLLM. Setup takes about 15 minutes, according to the lab's Korean-narrated guide.

What this means

KORMo-Web-11B targets teams that want one open model for text, image understanding and browser control, with Korean as a first-class language. Its Korean score of 30.1 is its strongest result in the lab's table, and its English score of 29.0 sits below OpenWebRL-4B. Because each agent in the comparison uses its own prompt and action format, the numbers show relative standing on this setup rather than a uniform harness. The Apache-2.0 license and published weights allow commercial deployment and fine-tuning, while the gap to the proprietary GPT-6-luna in the table remains large.

Related Articles

model release

KiteFishAI releases NanoI1, a 155-million-parameter open-weight model for local enterprise use

KiteFishAI has released NanoI1, an open-weight model with 155 million parameters. It targets specialized enterprise workflows on local hardware, according to Unite.AI.

model release

RadixArk releases DaisyRose-27B, lifting Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%

RadixArk released DaisyRose-27B, its first model, built on Qwen3.8-27B and trained with reinforcement learning on 1,200 curated agent tasks. According to RadixArk, it raises Terminal Bench 4.1 pass@1 from 5.6% to 12.0% and scores higher than the base model on all seven reported benchmarks. Weights ship under Apache 2.0 in BF16, FP8 and NVFP4.

model release

Google releases Genie 4 to AI Ultra subscribers with real-time prompt editing and native audio

Google is rolling out Genie 4, which it calls its most realistic and interactive world model yet, to eligible Google AI Ultra subscribers in Project Genie. The update adds real-time native audio, mid-session text prompting, a world-building agent and a community gallery. Genie 3 also gains sessions of up to 3 minutes.

model release

Microsoft releases Aion 1.0 Instruct, a 2.3B on-device SLM with 32K context under MIT license

Microsoft's Windows Applied Sciences Group has published Aion 1.0 Instruct, a roughly 2.29B-parameter instruction-tuned small language model with a 32,768-token context window under the MIT license. A further optimized version is slated to become the Windows OS SLM by the end of January 2027.

Comments

Loading...