Primary use stated or inferred cautiously from official documentation.
I
INCLUSIONAI / inclusionAI/LBS-27B
LBS-27B
inclusionAI vision model based on Qwen3.8-27b. It learns to locate interface elements and inspect image regions through cropping, transferring that strategy to general visual questions. The usable October 26 release includes twelve weight shards and a full model card; the empty repository earlier that morning was not considered a release. Its output consists of text and crop actions, not generated images.
01 / SUMMARY
The essential
Maximum input capacity when published by the source.
Documented maximum generation limit.
Verified availability channels.
Primary functional family and verified specialties.
Declared terms for API access, model weights, or self-hosting.
02 / TECHNICAL SHEET
Limits and integration
| API ID | inclusionAI/LBS-27B |
|---|---|
| Model type | Vision and documents · Multimodal |
| Access model | Open weights |
| License | Apache 2.0 |
| Deployment | Self-hosted with vLLM |
| Release | 26 OCT 2026 |
| Knowledge cutoff | Not published |
| Entrance | Text · Image |
| Exit | Text |
| Context window | Not published |
| maximum output | Not published |
| Reasoning | Not published |
| Published tools | crop_image |
| Structured Outings | Not published |
| Batch processing | Not published |
| Prompt cache | Not published |
| Fine-tuning | Not published |
| Verified platforms | Hugging Face · vLLM |
Locating controls, screenshots and high-resolution image perception. The card describes approximately 10,000 synthetic cold-start samples and 80 GRPO steps, without additional training on general VQA questions. It can request two independent crops per turn, with a total budget of three calls per episode in the published evaluation.
The client must execute crop_image and return images and coordinate mappings. Final coordinates are normalized to 0–1,000 on the original image. Locating an element does not execute an application action. No universally evaluated context or output limit has been verified: examples and configuration are not guarantees.
03 / CAPABILITIES
What it can do
Orientation
Locating controls, screenshots and high-resolution image perception. The card describes approximately 10,000 synthetic cold-start samples and 80 GRPO steps, without additional training on general VQA questions. It can request two independent crops per turn, with a total budget of three calls per episode in the published evaluation.
Context
Not published · Not published
Tools and integration
crop_image
Access
Hugging Face
Input modalities
TextImage
04 / PRICES
Documented cost
| Concept | Worth | Unit/condition |
|---|---|---|
| Standard input | No published official API rate | Standard API rate |
| Cached input | Not published | Reading reused prefixes |
| Cache write or storage | Not published | The condition varies by provider |
| Standard output | Self-hosted infrastructure cost | May include reasoning tokens |
| Batch input | Not published | Asynchronous processing |
| Batch output | Not published | Asynchronous processing |
The weights and project declare Apache 2.0; third-party datasets and screenshots retain their own terms. No official API price has been verified. The example uses four devices, without establishing a universal minimum memory requirement. Costs include inference, cropping and, where applicable, the judge model.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
05 / EVALUATIONS
How to read the results
“
Published zoom-in results: ScreenSpot-Pro 81.15% on 1,581 records, OS-World-G Refine 84.57% on 564 and PerceptionBench 59.67% on 3,000. PerceptionBench uses Deepseekv4-flash-0731 as judge; chart references use GPT-oss-120B. These different configurations prevent attributing a controlled improvement. Not reproduced by Inferama.
| Benchmark | Result | Metric | Source |
|---|---|---|---|
| ScreenSpot-Pro | 81,15 % | Action accuracy | View source ↗ |
| OS-World-G Refine | 84,57 % | Action accuracy | View source ↗ |
| PerceptionBench | 59,67 % | Judge accuracy | View source ↗ |
| HR-Bench 4K | 89,13 % | Strict matching | View source ↗ |
| HR-Bench 8K | 87,75 % | Strict matching | View source ↗ |
Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.
06 / VERSIONS
Chronology
LBS-27B
Version added to Inferama's verified catalog.
07 / SOURCES