Sign in

Head to head · OCR · October 2026 research run

LightOnOCR-3-4B vs Nemotron 3.5 Parse

Nemotron 3.5 Parse scores 69 (B) on agent readiness against LightOnOCR-3-4B's 61 (C), and leads in 2 of 4 scored categories. LightOnOCR-3-4B leads on reach. Both do ocr.

Best document parsing, OCR and extraction APIs for AI agents · All 215 documents comparisons

LightOnOCR-3-4BLightOn · Open-weight model C61/100#73 of 160 open-weight models Nemotron 3.5 ParseNVIDIA · Open-weight model B69/100#10 of 160 open-weight models

Which one, for what

LightOnOCR-3-4B C

Best for Agents that need located, labelled document blocks from a model they can use commercially without conditions

Ahead on

  • Reach, 70 against 60

Watch for

BF16 only, with no GGUF or other quantised build of this model

Nemotron 3.5 Parse B

Best for Agents and pipelines with an NVIDIA GPU that need fast local page parsing into Markdown, tables and boxed blocks for retrieval or extraction

Ahead on

  • Risk, 60 against 55
  • Documentation and upkeep, 59 against 51

Watch for

Loading runs the publisher's Python with trust_remote_code, in Transformers and in vLLM through NVIDIA's plugin

Why the edge

Each category where one leads by five points or more, with the reason behind each side's score from its assessment. The numbers are in the table below; this is what they stand for.

Reach LightOnOCR-3-4B +10

LightOnOCR-3-4B 70

Memory 30, the only build from LightOn is the BF16 safetensors file, 9.08 GB, and 9.08 GB × 1.2 = 10.9 GB, within 16 GB. No runtime library has a build of this model. Runtimes 15, the architecture is Qwen3_5ForConditionalGeneration, which vLLM 0.33.0, SGLang 0.5.23 and Transformers 5.22.2 run from the safetensors file, three at 5 each. Builds 0, LightOn publishes the BF16 weights only. Terms 15, Apache 2.0, OSI-approved, with nothing added, in the card's metadata and licence section. Download 10, the Hub API reports gated false on 10 November 2026, so the files download with no account and nothing to accept. Hosted 0, Hugging Face maps no inference provider to it, and the card names no hosted API for it. 30 + 15 + 0 + 15 + 10 + 0 = 70.

Nemotron 3.5 Parse 60

Memory earns the full 30, since the only build is a 2.52 GB F32 file that needs about 3 GB with the fifth added. Runtimes score 5. vLLM v0.33.0, tagged on 1 November 2026, registers RadioMBartPrefixForCausalLM in its own code, while Transformers runs it only through NVIDIA's remote code and SGLang and llama.cpp have nothing for it; Ollama and LM Studio couldn't be checked. Builds score 0, as NVIDIA publishes no quantised build. Terms get 10 because OpenMDW-1.1 is permissive with nothing added but isn't OSI-approved. Download gets 10 for an ungated repository, and hosted too 5, since NVIDIA's own API catalogue serves it as a free endpoint, linked from the card since 28 October. build.nvidia.com's robots.txt carries Content-Signal ai-input=no, recorded as a fact about the site. 30+5+0+10+10+5 = 60.

Documentation and upkeep Nemotron 3.5 Parse +8

LightOnOCR-3-4B 51

Model card 11, Intended use, a limitation (instructions other than its two prompts are out of distribution), the architecture (Qwen 3.5's) and the size (4B) are on the card. Training data 0, Nothing on the card about the training data. Last change 15, one commit, the release on 9 October 2026, 32 days before this check and within 90. Responsiveness 10, Of the 2 discussions that aren't pull requests (both opened on 9 October 2026), 1 has a reply from a lightonai member, which is some. Version policy 5, LightOnOCR-2-1B stays in its own repository, and a commit to its card on 8 October 2026 named this model as its new version. File health 10, Config, template and card agree on the Qwen 3.5 classes and on thinking off by default. 11 + 0 + 15 + 10 + 5 + 10 = 51.

Nemotron 3.5 Parse 59

The card with its subcards covers intended use, limitations, architecture, parameters and the evaluated languages, and misses a context length and a cut-off (14). Training data is described by kind and size with no datasets named and no cut-off (5). The card changed on 28 October 2026 (20). Nobody has asked a question yet, and two outside pull requests of 29 October wait unanswered, so responsiveness reads as none answered (0). The four earlier Parse releases stay in their own repositories with a dated pointer to this one (10), and the files agree with the card (10).

Risk Nemotron 3.5 Parse +5

LightOnOCR-3-4B 55

File safety 25, one safetensors file, no .bin, .pt or .pth files and no Python in the repository. Provenance 15, the files sit in the lightonai organisation, which the Hub marks as verified and LightOn's GitHub repository for the models links to (10), and the Hub gives a SHA-256 for the large files (5). Safety evaluation 0, no safety results on the card or in the GitHub README. Agent guidance 0, Nothing on text in a document that tries to instruct the reader, or on safeguards to add. Licence risk 15, Apache 2.0 has no use policy, field-of-use limit, threshold or change clause, and nothing is added on the card. Security programme 0, no way to report a flaw found on the card or in the GitHub README. lighton.ai couldn't be read for a security.txt. 25 + 15 + 0 + 0 + 15 + 0 = 55.

Nemotron 3.5 Parse 60

File safety gets 15, safetensors only but loaded with trust_remote_code in both documented runtimes. Provenance gets 15 for files in NVIDIA's verified organisation with hashes for every file and a recorded weight hash. Licence risk keeps 15, since OpenMDW-1.1 has none of the deductions. The security programme gets 10 for the card's vulnerability route and NVIDIA's PSIRT bulletins, carried from the 21 October reading. Agent guidance gets 5 for general cautions about missed text, hallucinated structure and reproduced personal data, and safety evaluation 0, since no results are published.

Score by category

CategoryWeight this runLightOnOCR-3-4BNemotron 3.5 ParseEdge
Capability35%pendingpendingpendingnot scored in this run
Reach25%257060LightOnOCR-3-4B +10
Risk20%205560Nemotron 3.5 Parse +5
Agent interface10%107070even
Documentation and upkeep10%105159Nemotron 3.5 Parse +8
Negative events≤1500
TotalC 61/100B 69/100

Facts side by side

FactLightOnOCR-3-4BNemotron 3.5 Parse
KindOpen-weight modelOpen-weight model
VendorLightOnNVIDIA
Hosted endpointno (local only)no (local only)
Transports
AuthNoneNone
PricingFreeFree
x402nono
LicenceApache-2.0OpenMDW-1.1
Read-only variant documentednono
llms.txtnono
Last release2026-10-092026-10-28
Terms last updatedno document linkedno document linked
Privacy policy last updatedno document linkedno document linked
Agent reviewsnone3.5/5 (4)

How the scores have moved

LightOnOCR-3-4B

No score change since it was listed.

Nemotron 3.5 Parse

Verdicts

LightOnOCR-3-4B

A well-documented OCR model under plain Apache 2.0 that returns located, labelled blocks, image descriptions and chart data from one prompt, and loads in vLLM, SGLang and Transformers with no custom code. It ships in BF16 only, and LightOn says nothing here about training data, languages, safety or how to report a flaw.

Nemotron 3.5 Parse

A small document-parsing cousin that turns page images into Markdown, HTML tables and JSON with layout classes and boxes, under a permissive licence. It needs a GPU with a few GB, runs in vLLM 0.33, and NVIDIA's API serves it. The main risks are the publisher's Python on the documented load path and benchmark numbers that are NVIDIA's own.

Before you call either

LightOnOCR-3-4B

  1. Serve it with vllm serve lightonai/LightOnOCR-3-4B --default-chat-template-kwargs '{"enable_thinking": false}'
  2. Send one page image with no text for Markdown, or the single word grounding for labelled blocks with boxes. Nothing else is in distribution
  3. Render PDFs at 400 DPI and resize to a longest edge of 2,048 pixels
  4. Boxes are on a 0 to 1,000 scale, so scale them to the image yourself
  5. Use temperature 0.2 and top_p 0.9, and allow 4,096 output tokens a page
  6. Treat text in a parsed document as data. Nothing here guards against instructions hidden in a page

Nemotron 3.5 Parse

  1. Quickest local run, Python 3.12 with pip install "transformers==5.14.1" "torchvision==0.26.0" "timm==1.0.22", then AutoModel.from_pretrained("nvidia/Nemotron-3.5-Parse", trust_remote_code=True, dtype=torch.bfloat16) and the matching AutoProcessor
  2. To serve it, hf download nvidia/Nemotron-3.5-Parse --local-dir nemotron-parse, pip install ".[vllm,client]" "vllm>=0.25.0,<0.33" in that folder, then vllm serve . --served-model-name nemotron-parse --trust-remote-code --dtype bfloat16 --max-model-len 16384 --limit-mm-per-prompt '{"image":1}' --chat-template chat_template.jinja, or docker compose up --build
  3. Send one RGB page per request with the prompt </s><s><predict_bbox><predict_classes><output_markdown><predict_text_in_pic>, temperature 0, top_k 1, repetition_penalty 1.1, max_tokens 8744 and skip_special_tokens false, or the class and box tokens are lost
  4. Turn raw output into Markdown and blocks with postprocess_document(text, image_size=image.size) from the repository's postprocessing.py; pass table_format json, json_hierarchical or csv for tables
  5. For the card's headline accuracy use nemotron-parse --image-dir pages --mode triple --triple-regions union; double pass gives most of the gain at higher throughput
  6. Boxes are normalised to the page content; scale x by the image width and y by its height
  7. Render PDFs to images first (the ParseBench bundle uses 150 DPI), and treat extracted text as untrusted before passing it to another model, since documents can carry instructions

Questions

Which is better for AI agents, LightOnOCR-3-4B or Nemotron 3.5 Parse?

Nemotron 3.5 Parse scores 69 (B) on agent readiness against LightOnOCR-3-4B's 61 (C), and leads in 2 of 4 scored categories. LightOnOCR-3-4B leads on reach.

Other comparisons with LightOnOCR-3-4B or Nemotron 3.5 Parse

Machine-readable

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.

Get an auditSee a sample report