PixWeb + PixVLM Small by PixLab In development

Small AI models for web automation, OCR and image understanding

Give browser agents a view of the page with PixWeb. Bring OCR, object recognition, captions and image-to-JSON extraction to your application with PixVLM Small. Both are being developed for local inference with WebGPU, WebAssembly and ONNX deployment targets.

Explore PixWeb Explore PixVLM Small Follow development

Development status: the capabilities below describe the product direction. Model downloads, browser demos, supported devices and measured benchmarks will be published with tested releases.

Browser page understanding

Screenshot → observation → planner/controller

PixWeb
Future tested PixWeb browser-understanding demonstration showing a source page and returned visual observation.

General image understanding

Image → answer, caption, OCR or selected fields

PixVLM Small
Future tested PixVLM Small demonstration showing an image beside an answer, caption, OCR result or selected structured fields.

PixWeb for web tasks

Understand screenshots, interface text, controls, forms and page state.

PixVLM Small for vision

Target object recognition, image questions, captions, OCR and field extraction.

Shared deployment direction

WebGPU, WASM CPU, ONNX and capable edge or embedded systems.

Choose by task

Two models. Two clear jobs

Choose the model by the task. PixWeb specializes in browser interfaces and observations for web automation. PixVLM Small covers general images, OCR, object recognition and extraction. Both share a browser and edge deployment direction; each will have its own weights, model card and measured results.

Choose PixWeb when

The source is a browser page

Use PixWeb when your application needs screenshot interpretation, UI text, form state, dialogs or visual observations before and after browser actions.

Choose PixVLM Small when

The source is a general image

Use PixVLM Small for photographs, receipts, labels and simple forms when you need OCR, object recognition, image Q&A, captions or selected fields.

Scroll horizontally to compare both models.

Comparison of PixWeb and PixVLM Small
Model Main purpose Example tasks Deployment targets

PixWeb

Web model

 Web understanding and support for browser automation. Screenshot interpretation, UI text, form states, dialogs and observations before or after browser actions. WebGPU, WASM CPU execution and compatible native ONNX runtimes.

PixVLM Small

Vision model

 General-purpose vision. OCR, object recognition, image Q&A, captions and image-to-JSON extraction. WebGPU, WASM CPU execution, native ONNX and capable embedded devices.

Separate training and evaluation for each model's tasks.

Shared integration patterns for browser and native applications.

Load only the model your application needs.

Artifact sizes and device measurements will be published with tested releases.

Explore PixWeb Explore PixVLM Small
PixWeb · Web-page understanding

Give your browser agent a view of the page

PixWeb focuses on what is visible in a browser: interface text, buttons, form fields, dialogs and page state. Turn a screenshot into an observation your application can use to understand a form, explain an error or decide what information the next step needs.

Read visible interface text

Read labels, validation messages, status text and the surrounding page context.

Describe controls, forms and layouts

Describe buttons, inputs, dashboards, forms and how visible elements relate to the page.

Surface blockers and uncertainty

Identify visible loading states, dialogs, workflow blockers and unclear or unreadable information for the integrating application.

Explore Browser Automation Developer Integration

PixWeb Page Understanding

Future real application screenshot placeholder

See

Visible UI

Understand

Page state

Return

Observation

PixWeb supplies visual page observations. The integrating application remains responsible for planning and browser actions.
PixWeb for AI browser automation

Observe the page. Take the next step. Check the result

Connect PixWeb to a planner and browser controller to support web-task automation. PixWeb supplies visual observations; the planner chooses a step, and the controller navigates, clicks, types or scrolls. A fresh screenshot helps the application check whether the page reached the expected state.

Browser Agent Workflow

Future real application screenshot placeholder

Placeholder for a future PixWeb browser automation integration demonstration.

  1. Inspect before an action

    Read visible fields, messages and available controls from the current screenshot.

  2. Provide workflow context

    Combine visual observations with DOM or accessibility information when your application supplies it.

  3. Let your application act

    The integrating planner and browser controller choose and execute navigation, clicks, typing or scrolling with their own permissions.

  4. Check the changed page

    Capture a fresh screenshot and inspect the resulting confirmation, error or changed state.

PixWeb observes. Your application acts.

PixWeb does not independently navigate, click, type, scroll or plan a browser task. The planner, browser access, controller and action permissions belong to the integrating application.

Explore web-agent integration

PixVLM Small Vision

Future real application screenshot placeholder

The same model is intended to cover general image understanding, local OCR and focused structured extraction.

Initial scope: object recognition does not automatically mean bounding-box detection, segmentation or tracking. Those are separate capabilities and are not promised by the initial scope.

PixVLM Small · General-purpose vision

Ask about the image in front of you

PixVLM Small focuses on everyday image understanding: recognize common objects, describe a scene, answer a visual question or produce a short caption. Use product photos, uploaded images and pictures captured by an application or device as the starting point for useful text.

Object recognition

Identify common visible objects and useful attributes.

Visual questions

Ask focused questions about visible image content.

Image captioning

Generate concise text for content and media workflows.

Scene understanding

Summarize what is visibly happening in an image.

Explore Local OCR Image-to-JSON Extraction
Local OCR

Read the text your application needs

Optical character recognition (OCR) turns text in an image into usable content. PixVLM Small targets local OCR for receipts, product labels, signs and simple forms. Ask for a total, read a printed identifier or extract a short passage close to the point of capture.

PixVLM Small OCR

Future real application screenshot placeholder

PixVLM Small is intended to read clearly visible text from receipts, labels, signs and simple forms.
Receipts
Merchant names, dates, totals and clearly visible fields.
Labels
Product text, printed identifiers and short instructions.
Image to text
Request a selected line or short passage.
Forms
Connect a visible field label with its visible value.

Focused OCR scope

Short, readable text is the initial target. Dense pages, handwriting and very small or blurry text need separate evaluation. Accuracy will be reported for tested model builds.

Turn Text into JSON Fields See OCR Use Cases
Image-to-JSON

Turn an image into the fields your workflow needs.

Ask for named values such as a receipt's merchant, date and total. Focused image-to-JSON extraction can connect receipts, labels and simple documents to an existing application without treating raw model output as already validated business data.

Merchant Date Total Selected fields

Structured extraction review flow

  1. Provide the source image

    Receipt, label or simple form with clearly visible text.

  2. Request named fields

    Keep the requested schema focused on values the workflow actually needs.

  3. Parse the response

    Check that the returned JSON shape matches what your application expects.

  4. Validate against the image

    Confirm important values and handle missing or unreadable information deliberately.

Web automation & vision AI use cases

Build around the pages and images your users already handle

Scroll horizontally to review all intended use cases.

Intended use cases for PixWeb and PixVLM Small
Workflow Model Intended contribution
Browser agents and web assistants PixWeb Page observations that help a planner choose and verify steps.
Form workflows and internal web tools PixWeb Visible fields, validation messages and completion states.
Browser testing and support tools PixWeb Screenshot summaries, errors and changed UI states.
Receipt capture and simple document entry PixVLM Small Local OCR and selected fields as JSON.
Product photos and content applications PixVLM Small Object recognition, image questions and captions.
Kiosks, gateways and embedded applications PixVLM Small General vision and image-to-text processing near the device.
Discuss Your Use Case See How the Workflows Fit
Two application workflows

Start with the task. Choose the model. Review the outcome

PixWeb workflow

A browser observation loop

  1. Capture the page

    Provide a screenshot and a focused question.

  2. Read the observation

    Inspect visible content, controls and page state.

  3. Take an action

    Let your planner and browser controller choose and execute the next step.

  4. Check the result

    Capture the updated page and compare it with the expected outcome.

PixVLM Small workflow

An image-to-result workflow

  1. Provide an image

    Add a photo, receipt, label or simple form.

  2. Ask for the output

    Request object information, a caption, OCR text or named fields.

  3. Review the response

    Compare it with the source image and validate extracted data.

  4. Integrate on your device

    Measure quality, loading, memory and response time for the intended workload.

Follow Model Development Developer Integration
Developer integration

Load a model. Supply an image. Use the result

The shared TypeScript integration is planned around one image or screenshot, a task prompt and a response with runtime information. Connect PixWeb observations to your browser agent workflow. Use PixVLM Small for image answers, captions, OCR text and selected fields. ONNX exports provide the corresponding route for native edge applications.

PixWeb TypeScript Example
// Planned integration shape — illustrative API

// 1. Load PixWeb with a local browser runtime
const model = await loadModel({
  model: "pixweb",
  backend: "webgpu"
});

// 2. Supply a browser screenshot and focused task
const result = await model.run({
  image: screenshot,
  prompt: "Describe the visible form state and any blockers."
});

// 3. Pass the observation to your application
console.log(result.text);

// Runtime information can identify the selected local backend
console.log(result.runtime);

// Example response shape
{
  "text": "A sign-in form is visible with email and password fields. The submit button appears disabled because the password field is empty.",
  "runtime": {
    "model": "pixweb",
    "backend": "webgpu"
  }
}

Illustrative integration shape only. Final API names, response fields and runtime interfaces will be published with the tested developer release.

Select the model for the job

Choose PixWeb for browser screenshots or PixVLM Small for general images, OCR and extraction.

Track loading and runtime state

Keep model loading progress and the selected local backend visible to the application.

Connect PixWeb to your browser agent

Pass observations to the planner and browser controller already provided by your application.

Validate PixVLM Small output

Check structured fields in application code and evaluate native ONNX builds on the device you intend to ship.

Follow Integration Development
Local browser inference

Bring visual inference into your browser application

Both models target WebGPU for browser GPU acceleration, with a separate WebAssembly (WASM) CPU path to validate for devices without WebGPU. The intended local inference path processes images or screenshots on the device when the required model and runtime assets are available locally.

WebGPU

The primary browser GPU-inference target for supported browsers and tested model profiles.

WASM CPU execution

A separate CPU execution target with its own compatibility, operator, memory and response-time checks.

Local assets & runtime visibility

Cache complete model assets for repeat use and identify the selected backend and model profile.

Explore Edge Deployment ONNX & Quantization

Browser inference architecture

Image or screenshot

PixWeb or PixVLM Small

Selected model assets + processor/tokenizer + local runtime

WebGPU

Browser GPU target

WASM CPU

Separately tested CPU target

Local model inference

Images or screenshots can remain on the device when the complete local path is used.

Optional connected components

A hosted planner or live website still has its own network and data flow.

Edge & embedded direction

Put image understanding close to the point of capture

PixVLM Small targets receipt reading, object recognition and image questions in kiosks, gateways and on-site applications. PixWeb can support browser-based interfaces on capable edge systems. Native ONNX deployment targets systems with the memory, compute and runtime compatibility needed for the selected build.

Deployment target, not universal device support
Discuss Embedded Deployment

Intended deployment environments

Kiosks & terminals

Receipts, labels, uploaded images and local vision workflows.

Gateways & edge computers

Image-to-text and general vision near the local device.

Embedded Linux

Compatible native runtimes on capable local systems.

Edge browser interfaces

PixWeb observations for browser UIs running on capable systems.

Support must be earned through testing

Memory

Measure the complete selected model build and runtime on the intended device.

Operators

Confirm that the chosen native runtime supports the exported model operations.

Compatibility

Validate the runtime, model profile and surrounding application together.

Response time

Measure the intended workload rather than assuming one device represents another.

ONNX · INT4/q4 · INT8

A deployment path for browsers and native applications

Open Neural Network Exchange (ONNX) is the export target for both models. Evaluate compact INT4/q4 and INT8 profiles alongside reference exports, then choose the tested build for your runtime. Model components may use different precisions, so download size, memory and quality are measured for the complete package.

Reference

Export the model

Create reference ONNX artifacts for compatible browser and native inference runtimes.

Evaluate

Compare precision profiles

Evaluate INT4/q4 and INT8 alongside reference exports. Model components may use different precisions.

Validate

Test the complete package

Measure download size, memory, quality, operator support and response time for each tested deployment profile.

ONNX artifacts

For compatible browser and native inference runtimes.

INT4/q4 & INT8

Evaluate storage and deployment tradeoffs.

Reference exports

Check conversion and quantization accuracy.

Separate validation

Validate WebGPU and CPU/WASM profiles independently.

Complete package

Model files, processor, tokenizer and required notices together.

Explore Developer Integration Follow Export Development
Commercial, embedded & OEM

Plan a deployment that fits your product

Talk to PixLab about integrating web understanding or general vision into a browser application, local tool or embedded product. Commercial terms depend on the selected model, applicable upstream requirements and the intended distribution.

Browser-agent integration OCR & vision applications Embedded and OEM requirements Distribution and release planning
Discuss Commercial Deployment View Development on GitHub

Licensing boundary

Repository source

The development repository's source-code license is AGPL-3.0-only.

Model artifacts

Weights, adapters and exported builds may have separate terms, including applicable upstream model licenses.

Upstream terms remain applicable

PixLab commercial terms do not override conditions attached to upstream models or components.

Release documentation

Released artifacts will identify their licenses, required notices, supported configurations and distribution conditions.

FAQ

Practical answers about PixWeb & PixVLM Small

Understand which model fits your task, what local inference means, how browser automation is divided between components, and what remains subject to testing before release.

Follow Development Compare Models
01

Models & capabilities

What are PixWeb and PixVLM Small?

PixWeb and PixVLM Small are PixLab's two small vision language models in development. PixWeb focuses on web pages, screenshots and observations for browser automation. PixVLM Small focuses on general image understanding, OCR, object recognition, captions and structured extraction.

Which model should I choose?

Choose PixWeb for browser agents, UI text, form states and web-task automation. Choose PixVLM Small for photographs, objects, receipts, labels, image questions and image-to-JSON extraction. Each model has its own task focus; an application does not need to load both.

How does PixWeb support web automation?

PixWeb is designed to describe a page's visible content, controls and state. A connected planner uses those observations to choose the next step, and a browser controller executes it. A fresh screenshot then helps the application check the result. See the browser automation workflow.

Does PixWeb click buttons or operate a browser by itself?

No. PixWeb supplies visual observations. The integrating application provides the planner, browser tools and action permissions that navigate, click, type and scroll. The model alone is not a complete browser agent.

Can PixWeb read text in screenshots?

Reading visible UI text is a core PixWeb target, including button labels, form fields, validation messages and status indicators. Its focus is web-page context. For receipts, product labels and general-image OCR, use PixVLM Small's OCR direction.

Can PixVLM Small perform OCR?

OCR is a core target for PixVLM Small: read short, clearly visible text in receipts, labels, signs and simple forms. Dense pages, handwriting and very small or blurry text require separate evaluation. Accuracy will be reported for tested model builds.

What does object recognition mean in PixVLM Small?

PixVLM Small targets identifying common objects and describing their visible attributes and relationships in an image. This supports image questions and captions. Detection boxes, segmentation and tracking are separate capabilities and are not promised by the initial scope.

Can PixVLM Small extract receipt data as JSON?

Receipt OCR and image-to-JSON extraction are intended workflows. Request named fields such as merchant, date and total, then validate the JSON structure and field values in your application. Missing or unreadable information should be surfaced for review.

02

Deployment & local processing

Will both models run in the browser with WebGPU?

WebGPU is the primary browser acceleration target for both models. Support will depend on the browser, GPU, available memory and released model profile. Each checkpoint will have its own tested configurations and measurements.

Will WebAssembly or WASM work without WebGPU?

A genuine CPU inference path through WebAssembly is a deployment target for both models. It must be validated separately from WebGPU for operator support, memory and response time. A WebGPU runtime containing a WASM file does not establish CPU fallback support.

Are ONNX, INT4 and INT8 models planned?

Yes. ONNX is the export target for browser and compatible native runtimes, with INT4/q4 and INT8 profiles evaluated alongside reference exports. A deployed model may combine different precisions across its components. Released profiles will identify actual download size, memory use and measured quality.

Can these models run on embedded devices?

Capable embedded devices, edge computers, kiosks and gateways are deployment targets through compatible native ONNX runtimes. PixVLM Small covers general vision and OCR; PixWeb can support browser-based interfaces on such systems. Device support will be listed after testing memory, operators and performance.

Can the models work offline?

Local inference can operate offline when the complete model and runtime assets are stored on the device and the application supports offline use. Live websites still need connectivity, and a hosted automation planner also needs a network connection. Offline model inference does not make the entire web workflow offline.

Will images and screenshots stay on the device?

The local inference path is intended to process images and screenshots on the device. For PixWeb, a connected planner may receive page observations or other application data according to its configuration. Applications should identify the inference route and any data sent to external services.

03

Development & licensing

How will developers integrate the models?

A shared TypeScript integration is planned for loading a selected model, supplying an image or screenshot and receiving text with runtime information. PixWeb observations connect to a planner and browser controller. PixVLM Small supplies answers, captions, OCR text or requested fields. ONNX exports provide the native deployment route.

Can I download or try the models now?

Both products are in development. Model downloads, browser demos, ONNX exports and benchmarks will be announced as they become available. The development repository currently documents the project direction.

Are the models open source?

The repository's source-code license is AGPL-3.0-only. Model weights, adapters and exported builds have separate terms, including applicable upstream model licenses. Each release will identify its permissions and required notices; source availability does not imply unrestricted use of every model artifact.

Can I use PixWeb or PixVLM Small commercially?

Commercial and OEM use depends on the chosen model, applicable upstream terms and the distribution arrangement. PixLab can discuss licensing, integration and support for its contributions. Contact [email protected] with your use case; PixLab terms do not remove upstream model-license conditions.

PixWeb + PixVLM Small · In development

Follow two small models built for two different vision jobs

PixWeb is being developed around browser-page observations for automation systems. PixVLM Small targets general image understanding, OCR and structured extraction. Tested releases will publish model artifacts, supported configurations and measured results as they become available.

Web understanding OCR & image understanding Local deployment direction
Follow Development on GitHub Discuss Commercial Deployment

No public production download or browser demo is being presented on this page while the models remain in development.