# LBS-27B (slim)

> LBS-27B is an open-weight vision-language model from inclusionAI's UI-Venus line, fine-tuned from Qwen3.8-27B to find what to click on a screenshot by zooming into regions over several turns, and applied to visual questions on large images.

- Full: https://www.anchorterminal.com/tools/lbs-27b.md (~9,500 tokens) · this version ~2,230 tokens · JSON https://www.anchorterminal.com/tools/lbs-27b.json · canonical https://www.anchorterminal.com/tools/lbs-27b
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-27 · pricing https://www.anchorterminal.com/tools/lbs-27b#pricing · reviews https://www.anchorterminal.com/tools/lbs-27b#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**D · 53.9/100 · rank #91 of 113 open-weight models · #73 in Open-weight models · not agent-ready · confidence low**


Assessment: A 27B Qwen3.8 fine-tune for GUI grounding that zooms into screenshots over a few turns before naming a point, and inclusionAI says it also helps with fine detail in large images. It needs about 96 GB at BF16 and a client that runs its crop actions. Every result so far is inclusionAI's own, with no safety results or injection guidance.

## Facts

- Kind: Open-weight model · vendor: inclusionAI (Ant Group) · category: Open-weight models · legal entity: inclusionAI, Ant Group (named in the site footer of inclusion-ai.org); the work is credited to authors from Wuhan University of Technology, CASIA, Zhejiang University and Ant Group · provenance 57/100
- Packages: huggingface `inclusionAI/LBS-27B`
- Auth: None · pricing: Free · x402: no · licence: Apache-2.0 (the card's metadata and licence section; the LICENSE file sits in the GitHub repository inclusionAI/UI-Venus on its LBS branch, not in the Hugging Face repository)
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/inclusionAI/LBS-27B, read from the card, the files and their metadata, the chat template, the LICENSE and READMEs on the LBS branch of github.com/inclusionAI/UI-Venus (by raw file) and the runtimes' own code on 27 October 2026. Nothing was downloaded or run
- Model: Qwen3_5ForConditionalGeneration (model_type qwen3_5), 27,356,728,560 BF16 parameters by the Hub's count. 64 text layers, three Gated DeltaNet linear-attention layers to each full-attention layer, hidden size 5,120, 24 attention heads and 4 KV heads, vocabulary 248,320, and a 27-layer vision encoder with 16-pixel patches. Images and video in, text out. No custom code and no auto_map
- Files: 54.73 GB in all, 12 BF16 safetensors shards, tokenizer.json, chat_template.jinja, configs for the model, generation, processor and preprocessor, and the card. No pickle files, no Python files and no LICENSE file in the Hugging Face repository. SHA-256 on every LFS file
- Licence: Apache 2.0. The card's metadata says apache-2.0, and its licence section says the weights and the project software are released under Apache 2.0, with third-party datasets and screenshots keeping their own terms. The LICENSE on the GitHub LBS branch is the Apache 2.0 text with the appendix left unfilled. The base, Qwen3.8-27B, is Apache 2.0, copyright Alibaba Cloud. No use policy found
- Runtimes: The card serves it with vLLM through the publisher's wrapper script. Released vLLM 0.32.0, SGLang 0.5.22 and Transformers 5.20.2 carry Qwen3_5ForConditionalGeneration, and llama.cpp b11888's converter maps it, with a qwen3vl vision projector. No Ollama library entry, no LM Studio catalogue entry and no ggml-org or lmstudio-community build on 27 October 2026
- Context: 262,144 tokens in config.json (max_position_embeddings, rope_type default), inherited from Qwen3.8-27B. The card states no context length, and the publisher's serving script defaults to 65,536 tokens and warns below 32,768 that multi-turn image context may overflow
- Tools and reasoning: The crop action is plain text in the reply, parsed by the publisher's own Python, not a runtime parser, and the system prompt that declares it is built by the publisher's build_system_prompt. The inherited Qwen template still declares tools in <tools> tags with <tool_call> blocks, which LBS's training didn't use. Thinking is on by default; reasoning_effort takes xhigh, medium or low, and enable_thinking false turns it off. The publisher's wrappers send reasoning_effort xhigh and a thinking_budget of 4,096 that the template doesn't read
- Hardware: Not stated in memory. The card's example uses four devices with tensor parallelism, and the training README says the recorded run used Ascend NPUs. The BF16 files need about 66 GB with a fifth added, so 96 GB of memory, within the 128 GB reference machine's GPU share
- Training data: Described in general terms, about 10,000 synthetic GUI grounding examples for supervised fine-tuning and 80 GRPO steps on synthetic GUI grounding tasks, with no general-domain data. No sources, sizes beyond that or cut-off. Eight training prompts and 32 rollouts from the run are in the GitHub repository

- Scores: Capability pending, Reach 55, Risk 50, Agent interface 70, Documentation and upkeep 43 · total over the 7 assessed categories
- Why:  · Reach, Memory earns 10, since the only build is 54.73 GB of BF16, about 66 GB with a fifth added, which fits the reference machine but not 64 GB. · Risk, File safety earns 15, safetensors only but with the documented route running the publisher's scripts to serve it and to write its system pro… · Agent interface, The chat template earns 20, Qwen's Jinja file in the repository. · Documentation and upkeep, The model card earns 8, with intended use, one limitation and parameters, and no architecture, context, languages or cut-off.
- Sources: 24, open questions: 10, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/lbs-27b.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/lbs-27b.svg` or a link to https://www.anchorterminal.com/tools/lbs-27b from a page on inclusion-ai.org or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Serve with `vllm serve inclusionAI/LBS-27B --tensor-parallel-size 4 --max-model-len 65536`, which is what the card's wrapper script runs; keep the context at 32,768 or more for multi-turn images
2. Build the system prompt with `build_system_prompt` from the LBS branch's `zoomin_grounding_protocol.py`, send the screenshot after `image_path: <name>`, and don't pass `tools`; the model answers in plain text
3. Parse `<action>crop_image(image_path=..., x1=..., y1=..., x2=..., y2=...)</action>`, cut the crop yourself and return it with its mapping to the original image; allow at most two crops per turn and three per episode
4. Read the final answer as `[x.xx,y.yy]` on a 0 to 1000 scale of the original image, and use your own executor to click
5. Pass `reasoning_effort` as `xhigh`, `medium` or `low` only, since the template raises an error for `high`, and don't count on `thinking_budget`, which the template ignores
6. Don't turn on MTP speculative decoding; the MTP weights aren't in the repository though config.json names one MTP layer

## Connect

```bash
# The card's route clones the LBS branch and wraps vllm serve (four GPUs in its example)
git clone --branch LBS --single-branch https://github.com/inclusionAI/UI-Venus.git
cd UI-Venus/evaluation
bash scripts/serve_vllm.sh start inclusionAI/LBS-27B --name lbs --tp 4 --dp 1 --daemon
# the same server without the wrapper (the script's defaults are a 65,536-token context and 64 sequences)
vllm serve inclusionAI/LBS-27B --served-model-name lbs --tensor-parallel-size 4 --max-model-len 65536
```

```bash
curl http://localhost:8000/v1/chat/completions -H 'Content-Type: application/json' \
  -d '{"model": "lbs", "max_tokens": 2560,
       "messages": [{"role": "system", "content": "<task, image size, crop_image spec and answer format, as build_system_prompt in the LBS branch writes them>"},
                    {"role": "user", "content": [{"type": "text", "text": "image_path: screenshot.png"},
                                                {"type": "image_url", "image_url": {"url": "data:image/png;base64,..."}}]}],
       "chat_template_kwargs": {"reasoning_effort": "xhigh"}}'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 74.7 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| Qwen3-Embedding-4B | BB | 72.9 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.min.md |
| Gemma 4 26B A4B IT | BB | 72.2 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-26b-a4b-it.min.md |
| Gemma 4 12B IT | BB | 71.3 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-12b-it.min.md |


## Panel reviews (0, desk reviews from public material, no calls made)