Skip to content
Model infoModelsSnapshot: 15 Oct 2026

Nanonets / Nanonets OCR

Nanonets OCR-3

Document-focused mixture-of-experts model for parsing, extraction, splitting, chunking and visual question answering.

Document extraction / OCR
availableClosed weightsDocument extractionOCR
Official documentation (opens in a new tab)Reviewed 15 Oct 2026
10 of 23 specification fields collected
Context windowNot collectedAwaiting source review
Maximum outputNot collectedAwaiting source review
Input billingNot collectedNanonets OCR-3 API / product
Output billingNot collectedNo inferred price

Specifications

Provider-documented facts. “Not collected” means we have not verified that field yet, not that the feature is unsupported.

Context / input limit
Not collected

Not established for this exact model. Workflow processing, output documents and marketing claims are not LLM capability or token-limit contracts.

Maximum input tokens
Not collected

Not established for this exact model. Workflow processing, output documents and marketing claims are not LLM capability or token-limit contracts.

Maximum output
Not collected

Not established for this exact model. Workflow processing, output documents and marketing claims are not LLM capability or token-limit contracts.

Maximum input content size
Not collected

Not established for this exact model. Workflow processing, output documents and marketing claims are not LLM capability or token-limit contracts.

Reasoning
Not collected

Not established for this exact model. Workflow processing, output documents and marketing claims are not LLM capability or token-limit contracts.

Tool calling
Not collected

Not established for this exact model. Workflow processing, output documents and marketing claims are not LLM capability or token-limit contracts.

Structured output
SupportedSource for Structured output (opens in a new tab)

Document /extract operation accepts a schema. Supported schema keywords and enforcement details are not established.

Streaming
Not collected

Not established for this exact model. Workflow processing, output documents and marketing claims are not LLM capability or token-limit contracts.

Batch processing
Not collected

Not established for this exact model. Workflow processing, output documents and marketing claims are not LLM capability or token-limit contracts.

Fine-tuning
Not collected

Not established for this exact model. Workflow processing, output documents and marketing claims are not LLM capability or token-limit contracts.

Context caching
Not collected

Not established for this exact model. Workflow processing, output documents and marketing claims are not LLM capability or token-limit contracts.

Parameters
35B total parameters reportedSource for Parameters (opens in a new tab)

Active parameter count is not given. Do not use total size as an active-compute figure.

Architecture
Mixture-of-experts vision-language model; selected experts run per token. The hosted API also integrates OCR engines for character-level extraction.Source for Architecture (opens in a new tab)
Languages
Not collected

Not established for this exact model. Workflow processing, output documents and marketing claims are not LLM capability or token-limit contracts.

Release date
2026-04-02Source for Release date (opens in a new tab)
Knowledge cutoff
Not collected

Not established for this exact model. Workflow processing, output documents and marketing claims are not LLM capability or token-limit contracts.

Input modalities
Document, Image, TextSource for Input modalities (opens in a new tab)
Output modalities
Text, JSON, Markdown, HTMLSource for Output modalities (opens in a new tab)
Output formats
Markdown, HTML tables, schema-shaped extraction and grounded answers with source boxes.Source for Output formats (opens in a new tab)
Tokens per image
1,280Source for Tokens per image (opens in a new tab)

Reported inference token cap per image. This is not a total prompt/context limit, output cap, or published billable token price.

Training dataset
More than 11 million documents reportedSource for Training dataset (opens in a new tab)

Provider-reported training scope, not an independent dataset audit.

Published training method
Low/high-resolution image curriculum; frozen backbone layers, EWC and replay examples to reduce forgetting.Source for Published training method (opens in a new tab)

These describe provider training, not a customer fine-tuning API.

Published price
Not collected

Exact OCR-3 operation pricing not established. Generic workflow block rates and older DocStrange free allowances are not assigned to this checkpoint.

Access & pricing

Prices belong to an endpoint and billing tier. A provider listing does not guarantee access for every account or region.

Nanonets OCR-3 API / product

Provider source (opens in a new tab)
API model ID
Not collected

Exact versioned API selector is not established.

Access
Official launch advertises a hosted API and directs users to the Document Parsing & Extraction product. Exact checkpoint selector and authenticated access were not verified.
Regions
Exact model-specific residency not established
API parameters
Published operations: /parse, /extract, /split, /chunk and /vqa. Exact versioned base URL and request schemas are not established.Source for API parameters (opens in a new tab)
Source locations and confidence
Per-element bounding boxes and confidence scores; visual answers can identify supporting page regions.Source for Source locations and confidence (opens in a new tab)

Endpoint pricing has not been collected. Check the provider for applicable rates.

Data & enterprise terms

No exact-version contractual retention, training-use or residency terms verified.

Caching

Caching can reuse earlier input processing. Availability, lifetime and billing depend on the endpoint. A cache hit still generates a new response.

01 / INITIAL INPUTProcess & retain

The system processes eligible input and retains reusable state.

02 / LATER REQUESTCheck available state

Reuse depends on matching input, session state and the cache policy.

03 / CACHE HITReuse & generate

Reusing state can reduce repeated work. Any charges follow the billing rules below.

Nanonets OCR-3 API / product

Not verified

No documented prompt-cache contract or discounted cache rate. Stored files, signed URLs and instant learning are not prompt caching.

How it works
Not collected

No documented prompt-cache contract or discounted cache rate. Stored files, signed URLs and instant learning are not prompt caching.

Minimum reusable input
Not collected

No documented prompt-cache contract or discounted cache rate. Stored files, signed URLs and instant learning are not prompt caching.

Lifetime & refresh
Not collected

No documented prompt-cache contract or discounted cache rate. Stored files, signed URLs and instant learning are not prompt caching.

How to enable it
Not collected

No documented prompt-cache contract or discounted cache rate. Stored files, signed URLs and instant learning are not prompt caching.

Usage fields
Not collected

No documented prompt-cache contract or discounted cache rate. Stored files, signed URLs and instant learning are not prompt caching.

Billing rules
Not collected

No documented prompt-cache contract or discounted cache rate. Stored files, signed URLs and instant learning are not prompt caching.

Limits & data boundaries
Not collected

No documented prompt-cache contract or discounted cache rate. Stored files, signed URLs and instant learning are not prompt caching.

No verified cache rate for this endpoint. An empty rate is not a zero charge.

Caching reviewed 15 Oct 2026

What the measurements say

olmOCR benchmark, original evaluation

Nanonets Research · provider-reported
87.4score

Original score shown in the provider table. Do not replace this with its corrected evaluation.

Evaluation source (opens in a new tab)

olmOCR benchmark, provider-corrected evaluation

Nanonets Research · provider-reported
93.1score

Provider correction after reviewing evaluator failures. Not directly interchangeable with uncorrected results from other models.

Evaluation source (opens in a new tab)

OmniDocBench v1.5

Nanonets Research · provider-reported
90.5score

Provider-reported release-page evaluation; no independently reproduced result in this review.

Evaluation source (opens in a new tab)

Latency and throughput depend on the host, request size, reasoning settings and load. There is no measured speed or global usage figure in this record.

The record behind the model

Why it is included

Classification inference: Nanonets describes a new 35B hosted OCR-3 model, distinct from its reviewed public OCR-s and OCR2 weight releases. No corresponding OCR-3 weights were found in the official public inventory. Exact licensing and any later weight release remain unverified.

Classification evidence (opens in a new tab)

Lifecycle dates

OCR-3 model release announced2 Apr 2026Source (opens in a new tab)

Sources reviewed