Skip to content
Model infoModelsSnapshot: 15 Oct 2026

Snowflake / Arctic-Extract

Arctic-Extract

Proprietary document model for extracting entities, lists and tables from scanned or digital documents through Snowflake AI_EXTRACT.

Document extraction / OCR
availableClosed weightsDocument extractionOCR
Official documentation (opens in a new tab)Reviewed 15 Oct 2026
17 of 29 specification fields collected
Context windowNot collectedAwaiting source review
Maximum outputNot collectedAwaiting source review
Input billingSee ratesSnowflake AI_EXTRACT
Output billingSee ratesPublished units and conditions below

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

No exact hosted model specification established. SQL execution and research-only limits do not imply a general chat API capability.

Maximum input tokens
Not collected

No exact hosted model specification established. SQL execution and research-only limits do not imply a general chat API capability.

Maximum output
Not collected

No exact hosted model specification established. SQL execution and research-only limits do not imply a general chat API capability.

Maximum input content size
Not collected

No exact hosted model specification established. SQL execution and research-only limits do not imply a general chat API capability.

Parameters
Not collected

No exact hosted model specification established. SQL execution and research-only limits do not imply a general chat API capability.

Reasoning
Not collected

No exact hosted model specification established. SQL execution and research-only limits do not imply a general chat API capability.

Tool calling
Not collected

No exact hosted model specification established. SQL execution and research-only limits do not imply a general chat API capability.

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

Constrained extraction schema: top-level object; string, string-list or table sub-objects. Not arbitrary JSON Schema support.

Streaming
Not collected

No exact hosted model specification established. SQL execution and research-only limits do not imply a general chat API capability.

Batch processing
Not collected

No exact hosted model specification established. SQL execution and research-only limits do not imply a general chat API capability.

Fine-tuning
SupportedSource for Fine-tuning (opens in a new tab)

Customer fine-tuning is supported in four documented regions; training limits differ from inference.

Context caching
Not collected

No exact hosted model specification established. SQL execution and research-only limits do not imply a general chat API capability.

Release date
Not collected

No exact hosted model specification established. SQL execution and research-only limits do not imply a general chat API capability.

Knowledge cutoff
Not collected

No exact hosted model specification established. SQL execution and research-only limits do not imply a general chat API capability.

Input modalities
Document, Image, TextSource for Input modalities (opens in a new tab)
Output modalities
JSON, TextSource for Output modalities (opens in a new tab)
Input file formats
View all 19 entries
  • PDF
  • PNG
  • PPTX
  • PPT
  • EML
  • DOC
  • DOCX
  • JPEG
  • JPG
  • HTM
  • HTML
  • TEXT
  • TXT
  • TIF
  • TIFF
  • BMP
  • GIF
  • WEBP
  • MD
Source for Input file formats (opens in a new tab)
Document file limit
Less than 100 MB; up to 125 pages at scale_factor 1.0Source for Document file limit (opens in a new tab)

The documented file-size error threshold is 104,857,600 bytes. Higher scale factors lower the page limit.

Document page limits
125 pages at scale 1; 62 at scale 2; 50 at scale 2.5; 31 at scale 4Source for Document page limits (opens in a new tab)

Page limit = floor(125 / scale_factor). Valid scale is 1.0 to 4.0.

Output limits by task
512 tokens per entity question; 4,096 tokens per table answerSource for Output limits by task (opens in a new tab)

These are task-specific answer limits, not one global response-token allowance.

Request limits
100 entity-equivalent questions per call; at most 10 table questions. Each table question consumes 10 entity equivalents.Source for Request limits (opens in a new tab)
Output formats
JSON with extracted entities, lists and column-oriented tables; optional extraction confidence.Source for Output formats (opens in a new tab)
Languages
View all 29 entries
  • Arabic
  • Bengali
  • Burmese
  • Cebuano
  • Chinese
  • Czech
  • Dutch
  • English
  • French
  • German
  • Hebrew
  • Hindi
  • Indonesian
  • Italian
  • Japanese
  • Khmer
  • Korean
  • Lao
  • Malay
  • Persian
  • Polish
  • Portuguese
  • Russian
  • Spanish
  • Tagalog
  • Thai
  • Turkish
  • Urdu
  • Vietnamese
Source for Languages (opens in a new tab)
Architecture
Qwen2.5-VL architecture with 4-to-1 vision-token merging; provider-trained LoRA merged into the model and quantized with 4-bit AWQ.Source for Architecture (opens in a new tab)

Research configuration, not a claim about every hosted revision.

Research context window
128,000 tokens in the published research configurationSource for Research context window (opens in a new tab)

The hosted SQL interface exposes document/page limits rather than this general token budget.

Training dataset
372,544 training examples from 35 datasets, including internal and public-derived datasets.Source for Training dataset (opens in a new tab)

Provider-reported research training, not a data-rights audit.

Published training method
LoRA rank 32, alpha 32; AWQ group size 128 and GEMM quantization. Loaded research model occupies 6.6 GiB.Source for Published training method (opens in a new tab)

Memory footprint is not parameter count or a customer self-hosting license.

Usage metering
Page tokens + prompt tokens + output tokens. Document page baseline: 970 tokens, multiplied by scale_factor.Source for Usage metering (opens in a new tab)

970-token rule explicitly covers PDF, DOCX, TIF/TIFF pages and JPEG/JPG/PNG images. Other listed formats need confirmed billing conversion.

Published price
Snowflake AI Credits; contract-specific currency conversionSource for Published price (opens in a new tab)

Compute, storage and other platform charges can be additional.

Access & pricing

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

Snowflake AI_EXTRACT

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

No public checkpoint selector: the base function is implicit, while custom model builds use an account-specific registry name.

Access
Requires SNOWFLAKE.CORTEX_USER and access to the input text or staged file. Base Arctic-Extract is selected by the AI_EXTRACT function, not a user-supplied checkpoint ID. Model authorization also applies: model RBAC is checked first, then CORTEX_MODELS_ALLOWLIST; if neither permits access the call fails.
Regions
AWS: N. Virginia, Oregon, Canada Central, Sao Paulo, Ireland, Frankfurt, Tokyo, Sydney. Azure: East US 2, West US 2, South Central US, North Europe, West Europe, Southeast Asia, Australia East, Central India, Japan East. GCP: Middle East Central2 (Dammam). Cross-region routing can extend access; unavailable in PR China.
API endpoint URL
SQL AI_EXTRACT(text => ..., responseFormat => ..., scores => TRUE/FALSE) or AI_EXTRACT(file => ..., responseFormat => ..., config => ..., scores => TRUE/FALSE)Source for API endpoint URL (opens in a new tab)
API parameters
Use text or file, not both. Optional config.scale_factor ranges from 1.0 to 4.0 and defaults to 1.0. scores defaults to FALSE.Source for API parameters (opens in a new tab)
Source locations and confidence
Confidence values between 0 and 1; per-entity scores and aggregate list/table scores. No per-cell or per-list-item scores. No extra score charge.Source for Source locations and confidence (opens in a new tab)
Deployment
Snowflake-managed compute; a warehouse above MEDIUM does not improve AI_EXTRACT inference performance.Source for Deployment (opens in a new tab)
dataResidency
CORTEX_ENABLED_CROSS_REGION controls routing. Input/output payloads are transmitted transiently to processing regions and are not persisted there, according to Snowflake.Source for dataResidency (opens in a new tab)
Stage security requirements
Inference supports client-side and server-side encrypted stages, including PrivateLink/network-policy accounts. Training has different restrictions.Source for Stage security requirements (opens in a new tab)
Authentication
In addition to function/object privileges, model access must be granted via model RBAC or the model/alias must be permitted by CORTEX_MODELS_ALLOWLIST. All allows fallback access; otherwise an unpermitted call fails.Source for Authentication (opens in a new tab)

The base AI_EXTRACT governance alias is arctic-extract. Account-specific custom model privileges still apply.

Prices retain their published currencies and billing units below.

Published billing units · original currency · observed 15 Oct 2026
CategoryUsageRateUnit
Inference AI CreditsBase extraction · AI Credits15.55 credits1M processed tokens

Rate conditions

  1. Snowflake AI Credits, not USD. Contract credit conversion, warehouse/storage and other platform charges are separate. Page and prompt/output token accounting follows the relevant endpoint documentation.
Pricing source (opens in a new tab)

Data & enterprise terms

Snowflake states that customer data is not used to train models made available to its customer base. Except as elected, AI models run inside its security/governance perimeter. Cross-region processing can transiently transmit prompts/results outside the account region; exact contractual retention terms are not established here.

Read the service policy (opens in a new tab)

Snowflake AI_EXTRACT fine-tuned

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

No public checkpoint selector: the base function is implicit, while custom model builds use an account-specific registry name.

Access
Uses a customer-created Snowflake Model Registry model and the documented model privileges. No separate public foundation checkpoint is inferred. Model authorization also applies: model RBAC is checked first, then CORTEX_MODELS_ALLOWLIST; if neither permits access the call fails.
Regions
Validate the custom build and current regional availability. Training is documented in AWS Oregon, N. Virginia, Frankfurt and Azure East US 2.
API endpoint URL
AI_EXTRACT(model => 'database.schema.model_name', file => TO_FILE(...), config => {...})Source for API endpoint URL (opens in a new tab)
API parameters
The model argument identifies a customer registry object. Optional responseFormat can override the training questions; optional scores/config follow AI_EXTRACT support.Source for API parameters (opens in a new tab)
Authentication
In addition to function/object privileges, model access must be granted via model RBAC or the model/alias must be permitted by CORTEX_MODELS_ALLOWLIST. All allows fallback access; otherwise an unpermitted call fails.Source for Authentication (opens in a new tab)

The base AI_EXTRACT governance alias is arctic-extract. Account-specific custom model privileges still apply.

Prices retain their published currencies and billing units below.

Published billing units · original currency · observed 15 Oct 2026
CategoryUsageRateUnit
Inference AI CreditsFine-tuned extraction · AI Credits19.14 credits1M processed tokens

Rate conditions

  1. Snowflake AI Credits, not USD. Contract credit conversion, warehouse/storage and other platform charges are separate. Page and prompt/output token accounting follows the relevant endpoint documentation.
Pricing source (opens in a new tab)

Data & enterprise terms

Snowflake states that customer data is not used to train models made available to its customer base. Except as elected, AI models run inside its security/governance perimeter. Cross-region processing can transiently transmit prompts/results outside the account region; exact contractual retention terms are not established here.

Read the service policy (opens in a new tab)

Snowflake Cortex fine-tuning

Provider source (opens in a new tab)
API model ID
arctic-extract

Base-model identifier for FINETUNE CREATE only; not a base AI_EXTRACT model argument.

Access
Requires Dataset, stage, schema and CREATE MODEL privileges. Produces a custom registry model.
Regions
AWS US West 2 (Oregon), US East 1 (N. Virginia), Europe Central 1 (Frankfurt); Azure East US 2 (Virginia).
API endpoint URL
SNOWFLAKE.CORTEX.FINETUNE('CREATE', '@database.schema.model_name', 'arctic-extract', 'training_dataset', ...)Source for API endpoint URL (opens in a new tab)

Syntax block uses an @-prefixed model target; a later guide example uses an unprefixed name. Verify the target form against the current SQL reference when executing.

API parameters
max_epochs accepts integers 2 through 10; otherwise determined automatically. Optional validation Dataset.Source for API parameters (opens in a new tab)
Training limits
At most 1,000 unique documents. Questions × total referenced pages must be ≤ 50,000. Repeated file references count toward total pages. At least 20 documents recommended, not required.Source for Training limits (opens in a new tab)
Document page limits
Training: 64 pages per file in Oregon/Frankfurt; 125 in N. Virginia/Azure East US 2.Source for Document page limits (opens in a new tab)
Input file formats
PDF, PNG, JPG, JPEG, TIFF, TIFSource for Input file formats (opens in a new tab)
Stage security requirements
Training supports internal and named external stages, but not client-side encrypted stages. Currently incompatible with custom network policies.Source for Stage security requirements (opens in a new tab)

Prices retain their published currencies and billing units below.

Published billing units · original currency · observed 15 Oct 2026
CategoryUsageRateUnit
Fine-tuning Platform CreditsFine-tuning · Platform Credits10 credits1M processed tokens

Rate conditions

  1. Platform Credit training rate only; zero in this table does not waive warehouse, storage or other charges. No USD conversion inferred.
Pricing source (opens in a new tab)

Data & enterprise terms

Snowflake states that customer data is not used to train models made available to its customer base. Except as elected, AI models run inside its security/governance perimeter. Cross-region processing can transiently transmit prompts/results outside the account region; exact contractual retention terms are not established here.

Read the service policy (opens in a new tab)

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.

Snowflake AI_EXTRACT

Not verified

AI_EXTRACT prompt caching and discounted cached-token rates are not established.

How it works
Not collected

AI_EXTRACT prompt caching and discounted cached-token rates are not established.

Minimum reusable input
Not collected

AI_EXTRACT prompt caching and discounted cached-token rates are not established.

Lifetime & refresh
Not collected

AI_EXTRACT prompt caching and discounted cached-token rates are not established.

How to enable it
Not collected

AI_EXTRACT prompt caching and discounted cached-token rates are not established.

Usage fields
Not collected

AI_EXTRACT prompt caching and discounted cached-token rates are not established.

Billing rules
Not collected

AI_EXTRACT prompt caching and discounted cached-token rates are not established.

Limits & data boundaries
Published automatic response caching covers text-only AI_COMPLETE, not AI_EXTRACT. Its 24-hour TTL and free cache hits must not be applied to document extraction.Source for Limits & data boundaries (opens in a new tab)

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

Caching reviewed 15 Oct 2026

Snowflake AI_EXTRACT fine-tuned

Not verified

AI_EXTRACT prompt caching and discounted cached-token rates are not established.

How it works
Not collected

AI_EXTRACT prompt caching and discounted cached-token rates are not established.

Minimum reusable input
Not collected

AI_EXTRACT prompt caching and discounted cached-token rates are not established.

Lifetime & refresh
Not collected

AI_EXTRACT prompt caching and discounted cached-token rates are not established.

How to enable it
Not collected

AI_EXTRACT prompt caching and discounted cached-token rates are not established.

Usage fields
Not collected

AI_EXTRACT prompt caching and discounted cached-token rates are not established.

Billing rules
Not collected

AI_EXTRACT prompt caching and discounted cached-token rates are not established.

Limits & data boundaries
Published automatic response caching covers text-only AI_COMPLETE, not AI_EXTRACT. Its 24-hour TTL and free cache hits must not be applied to document extraction.Source for Limits & data boundaries (opens in a new tab)

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

Caching reviewed 15 Oct 2026

Snowflake Cortex fine-tuning

arctic-extract
Not verified

AI_EXTRACT prompt caching and discounted cached-token rates are not established.

How it works
Not collected

AI_EXTRACT prompt caching and discounted cached-token rates are not established.

Minimum reusable input
Not collected

AI_EXTRACT prompt caching and discounted cached-token rates are not established.

Lifetime & refresh
Not collected

AI_EXTRACT prompt caching and discounted cached-token rates are not established.

How to enable it
Not collected

AI_EXTRACT prompt caching and discounted cached-token rates are not established.

Usage fields
Not collected

AI_EXTRACT prompt caching and discounted cached-token rates are not established.

Billing rules
Not collected

AI_EXTRACT prompt caching and discounted cached-token rates are not established.

Limits & data boundaries
Published automatic response caching covers text-only AI_COMPLETE, not AI_EXTRACT. Its 24-hour TTL and free cache hits must not be applied to document extraction.Source for Limits & data boundaries (opens in a new tab)

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

Caching reviewed 15 Oct 2026

What the measurements say

No benchmark observations have been added for this model. We do not turn a vendor description into a performance score.

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

Snowflake explicitly calls Arctic-Extract a proprietary vision-based LLM. The research model uses the Qwen2.5-VL architecture; this does not make the provider-specific trained derivative a public-weight release.

Classification evidence (opens in a new tab)

Lifecycle dates

Customer fine-tuning became generally available20 Apr 2026Source (opens in a new tab)

Sources reviewed