The Open Weights
LatestModelsLeaderboardsCompanies
Subscribe
The Open Weights

The daily record of open-source AI. New model releases, leaderboards, and what's coming next — written for people who ship.

Refreshed every 12 hours

Discover

  • Latest releases
  • New today
  • Trending models

Browse

  • All models
  • Companies
  • Categories
  • Leaderboards

About

  • About
  • Editorial policy
  • Hardware estimates
  • RSS feed
  • llms.txt
  • Newsletter

© 2026 The Open Weights. An independent publication.

PrivacyTermsSMSAggregated by Claude · curated by humans.
LatestLightOn3-4B
LightOnVision-Language

LightOn ships a 4B OCR model with visual grounding

LightOnOCR-3-4B pairs document transcription with the ability to point back to where text sits on the page.

Oct 6, 2026
NotableOther
LightOnOCR-3-4B

LightOn has released LightOnOCR-3-4B, a 4-billion-parameter vision-language model aimed squarely at document understanding. Rather than a general-purpose chat model, it is tuned for the unglamorous but high-value work of turning pages—scans, PDFs, forms—into structured, machine-readable text.

The headline capability is visual grounding: beyond transcribing content, the model can tie its outputs back to specific regions of the source image. That matters for anyone who needs to trust and verify an extraction, since a grounded result can be audited against the original layout instead of taken on faith.

Why it matters

OCR has quietly become one of the most practical uses of multimodal models, and the trend is toward smaller systems that are cheap enough to run at scale.

  • At 4B parameters, the model is small enough to deploy on modest hardware while still handling complex layouts.
  • Grounding makes it a better fit for compliance, search, and data-pipeline work where provenance is required.
  • It arrives under an open weights release on Hugging Face, lowering the barrier for teams to self-host rather than route sensitive documents through an API.

The model is distributed under a custom license, so teams should check the terms before commercial deployment. As with any OCR system, real-world accuracy will depend heavily on document quality and language coverage, which prospective users will want to validate on their own corpora.

Sources

  • lightonai/LightOnOCR-3-4B

    Hugging Face

    Visit
OlderFalcon-Emirati tunes an LLM for local dialectNVIDIA · Text / LLM · last monthNewerMiniCPM-V 4.7 brings a 35B MoE vision model on-deviceOpenBMB · Vision-Language · last month

Get the model

Hugging Face

Specs

Parameters4B
Context window262K tokens
Size9.1 GB
PrecisionBF16
ArchitectureQwen3_5ForConditionalGeneration
LicenseOTHER
Downloads14.6K
Likes145

Can you run it?

Runs on a laptop — about 4.8 GB at 4-bit.

  • BF16 (as published)

    12 GB GPU (RTX 3060 / 4070) · 16 GB Mac

    11.2 GB
  • 8-bit

    8 GB GPU (RTX 4060) · 16 GB laptop or Mac

    6.9 GB
  • 4-bit

    8 GB GPU (RTX 4060) · 16 GB laptop or Mac

    4.8 GB
Your machine
GGUF builds
  • KV cache sized for 8,192 tokens of context; longer prompts need proportionally more.

Based on Hugging Face weights metadata. Assumes an 8K context and ~1 GB runtime overhead; actual needs vary. How we estimate


Modalities

Vision-Language

The Weekly Weights

Every open release that mattered, one email a week.

0 comments

No comments yet. Be the first to weigh in.

More from LightOn

All LightOn releases →
LightOnOCR-2 1B
LightOn/Vision-LanguageRuns anywhere

LightOn Releases OCR-2, a 1B Document AI Model

The new vision model from the Paris-based AI lab uses Mistral architecture to extract text and structure from complex documents like PDFs and forms.

Jan 16, 2026

More in Vision-Language

All Vision-Language →
Mirage
Cactus Compute/Vision-Language

Cactus Compute's Mirage brings vision tasks on-device

The new open vision-language model targets detection, segmentation, and visual Q&A without relying on the cloud.

Nov 5, 2026
Inkling-2-Small
Thinkingmachines/Vision-LanguageMulti-GPU server

Thinking Machines ships Inkling-2-Small, an open MoE VLM

The Apache-2.0 vision-language model uses a mixture-of-experts design and lands openly on Hugging Face.

Nov 4, 2026
LongCat 2.5
Meituan/Text / LLMDatacenter

Meituan Releases LongCat 2.5, an MIT-Licensed MoE

The Chinese delivery giant expands its open-model push with a multimodal, agentic mixture-of-experts system.

Nov 4, 2026