Skip to content
Model infoModelsSnapshot: 15 Oct 2026

Maritaca AI / Sabiá

Sabiazinho 4

Lower-cost Portuguese language model for high-volume extraction, classification, assistance and tool-driven workflows.

Language models (LLMs)
availableClosed weightsText generation
Official documentation (opens in a new tab)Reviewed 14 Oct 2026
18 of 22 specification fields collected
Context window128K128,000 tokens
Maximum output32.8K32,768 tokens
Input / 1M tokensR$1.00Maritaca API
Output / 1M tokensR$4.00BRL · billing 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
128,000 tokensSource for Context / input limit (opens in a new tab)

Exact numeric context from the documented model-list response, corroborated by the model catalog. Shared context, not an additional independent input allowance.

Maximum input tokens
Not collected
Maximum output
32,768 tokensSource for Maximum output (opens in a new tab)

Exact max_completion_tokens in the provider response example; the overview rounds this to 32K.

Input modalities
TextSource for Input modalities (opens in a new tab)

File and image input is converted to text by OCR. The underlying language model does not receive the original image.

Output modalities
TextSource for Output modalities (opens in a new tab)
Reasoning
Not collected
Tool calling
SupportedSource for Tool calling (opens in a new tab)

Function calling is documented for the Sabiá 4 family; integrated server-side tools have separate model and streaming restrictions.

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

Provider-wide Responses API JSON-schema support, with examples using sabia-4. Subset only: up to 100 object properties, five nesting levels, 15,000 total schema-name characters and 500 enum values. Not all JSON Schema constraints are supported.

Streaming
SupportedSource for Streaming (opens in a new tab)

Text streaming is supported. Integrated web search, code execution and Data Ocean require stream:false; asynchronous Batch has no streaming.

Batch processing
SupportedSource for Batch processing (opens in a new tab)

Rates published for this model and BR-SP. One model per JSONL file, at most 50,000 requests / 200 MB, 24-hour completion window; completed output files are kept for 30 days. Expiration is possible.

Fine-tuning
Not supportedSource for Fine-tuning (opens in a new tab)

Provider states customer fine-tuning is not offered currently and recommends RAG for customization.

Knowledge cutoff
Not collected
Release date
2026-01-05Source for Release date (opens in a new tab)

Public preview announcement date; the current catalog lists this model as available. Snapshot suffixes do not establish the public launch date.

Parameters
Not disclosedSource for Parameters (opens in a new tab)

The provider explicitly does not disclose exact parameter counts.

Architecture
TransformerSource for Architecture (opens in a new tab)

Exact base checkpoint and dense/MoE configuration are not identified. The training blog’s trillion-parameter cost exercise is illustrative, not this model’s specification.

Languages
PortugueseSource for Languages (opens in a new tab)

Documented specialization, not a claim that other languages are unsupported.

Context caching
SupportedSource for Context caching (opens in a new tab)

Automatic shared-prefix reuse; minimum eligible length and retention duration are not specified.

API aliases
sabiazinho-4, sabiazinho-4-2026-01-06, sabia-4-small, sabiazim-4Source for API aliases (opens in a new tab)
Published price
Brazilian reais (BRL), per 1M tokens; standard, night, Flex, Batch and cached-input rates vary by endpoint.Source for Published price (opens in a new tab)
fileProcessing
OCR text extraction from PDFs and PNG/JPEG images; extraction charges can be additionalSource for fileProcessing (opens in a new tab)
integratedTools
Web search, Page readingSource for integratedTools (opens in a new tab)

Server-side agent workflow returns only its final response and optional generated artifacts. Intermediate reasoning/tool calls are not exposed.

trainingPipeline
Portuguese and Brazilian legal continued pretraining, context extension, instruction fine-tuning, preference alignmentSource for trainingPipeline (opens in a new tab)

Published for Sabiá 4 and Sabiazinho 4. This recipe is not assumed to describe every additional Thinking-stage training step.

Access & pricing

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

Maritaca API

Provider source (opens in a new tab)
API model ID
sabiazinho-4
Access
API key required; Responses and Chat Completions. Enterprise terms available by agreement.
Regions
Standard endpoint; no Brazil-only residency guarantee. Provider uses several infrastructure partners.
apiBase
https://chat.maritaca.ai/apiSource for apiBase (opens in a new tab)
Equivalent API aliases
sabiazinho-4, sabiazinho-4-2026-01-06, sabia-4-small, sabiazim-4Source for Equivalent API aliases (opens in a new tab)
Batch limits
50,000 requests / 200 MB; one model per file; 24-hour completion windowSource for Batch limits (opens in a new tab)
flex
service_tier: flex; capacity-dependent, up to five-minute queue, then 429; immediate 429 if queue is fullSource for flex (opens in a new tab)

Prices retain their published currencies and billing units below.

Published billing units · original currency · observed 14 Oct 2026
CategoryUsageRateUnit
StandardInput1R$1.001M tokens
StandardOutput1R$4.001M tokens
CachingCached input1R$0.251M tokens
NightInput1R$0.701M tokens
NightOutput1R$2.801M tokens
FlexInput1R$0.501M tokens
FlexOutput1R$2.001M tokens
BatchInput1R$0.501M tokens
BatchOutput1R$2.001M tokens
Tool add-onWeb search2R$0.0165search
Tool add-onPage reading2R$0.066page read
File extraction add-onPDF extraction / medium3R$0.02PDF page
File extraction add-onPDF extraction / advanced3R$0.045PDF page

Rate conditions

  1. BRL per million tokens. Cache hits retain their own 75% input discount. Batch, Flex and night rates apply only to uncached input and output and are mutually exclusive. Night hours are 22:00–06:00 BRT. Flex is capacity-dependent and may queue up to five minutes or return 429. No currency conversion is applied.
  2. Additional server-side tool fee, separate from token charges. Integrated tools require stream:false. A web-search request may perform both searches and page reads. Data Ocean can also trigger search, reading and code fees. No BR-SP tool-fee premium is specified.
  3. Additional charge per PDF page, separate from model tokens. Repeated extraction can be billed again; keep the returned extracted text for subsequent requests. The level setting does not apply to other file types.
Pricing source (opens in a new tab)Pricing source (opens in a new tab)

Data & enterprise terms

Provider says API inputs/outputs are not used for model training. Normal content is discarded after generation, but error/observability logs may retain request content up to 30 days and infrastructure logs up to 18 months. Batch output files have a separate 30-day retention period; prompt-cache lifetime is unspecified. Chat/app conversations have different storage rules.

Read the service policy (opens in a new tab)

Maritaca API / São Paulo

Provider source (opens in a new tab)
API model ID
sabiazinho-4-br-sp
Access
API key required; Responses and Chat Completions. Enterprise terms available by agreement.
Regions
São Paulo, Brazil; inference, logs and telemetry in EVEO infrastructure, according to provider.
apiBase
https://chat.maritaca.ai/apiSource for apiBase (opens in a new tab)
Equivalent API aliases
sabiazinho-4-br-spSource for Equivalent API aliases (opens in a new tab)
Batch limits
50,000 requests / 200 MB; one model per file; 24-hour completion windowSource for Batch limits (opens in a new tab)
flex
service_tier: flex; capacity-dependent, up to five-minute queue, then 429; immediate 429 if queue is fullSource for flex (opens in a new tab)

Prices retain their published currencies and billing units below.

Published billing units · original currency · observed 14 Oct 2026
CategoryUsageRateUnit
StandardInput1R$1.301M tokens
StandardOutput1R$5.201M tokens
CachingCached input1R$0.331M tokens
NightInput1R$0.911M tokens
NightOutput1R$3.641M tokens
FlexInput1R$0.651M tokens
FlexOutput1R$2.601M tokens
BatchInput1R$0.651M tokens
BatchOutput1R$2.601M tokens
Tool add-onWeb search2R$0.0165search
Tool add-onPage reading2R$0.066page read
File extraction add-onPDF extraction / medium3R$0.02PDF page
File extraction add-onPDF extraction / advanced3R$0.045PDF page

Rate conditions

  1. BRL per million tokens. Cache hits retain their own 75% input discount. Batch, Flex and night rates apply only to uncached input and output and are mutually exclusive. Night hours are 22:00–06:00 BRT. Flex is capacity-dependent and may queue up to five minutes or return 429. No currency conversion is applied. BR-SP token charges include the published 30% premium; cached-input figures are the rounded published values, not an inferred higher-precision rate.
  2. Additional server-side tool fee, separate from token charges. Integrated tools require stream:false. A web-search request may perform both searches and page reads. Data Ocean can also trigger search, reading and code fees. No BR-SP tool-fee premium is specified.
  3. Additional charge per PDF page, separate from model tokens. Repeated extraction can be billed again; keep the returned extracted text for subsequent requests. The level setting does not apply to other file types.
Pricing source (opens in a new tab)Pricing source (opens in a new tab)

Data & enterprise terms

Provider says API inputs/outputs are not used for model training. Normal content is discarded after generation, but error/observability logs may retain request content up to 30 days and infrastructure logs up to 18 months. Batch output files have a separate 30-day retention period; prompt-cache lifetime is unspecified. Chat/app conversations have different storage rules.

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.

Maritaca API

sabiazinho-4
Documented
Support evidence (opens in a new tab)
How it works
Automatic prompt-prefix cacheSource for How it works (opens in a new tab)
Minimum reusable input
Not collected

Minimum eligible prefix length is not specified.

Lifetime & refresh
Not collected

No retention/TTL or refresh policy is published in the reviewed cache guide.

How to enable it
No explicit cache creation; keep stable content first and reuse identical token prefixes.Source for How to enable it (opens in a new tab)
Usage fields
prompt_tokens_details.cached_tokensSource for Usage fields (opens in a new tab)
Billing rules
Cached input costs 25% of standard input; discounts on uncached input/output do not compound with cache.Source for Billing rules (opens in a new tab)
Limits & data boundaries
Batch, Flex and night pricing are mutually exclusive. Cache scope, cross-account isolation and regional reuse are not specified in the reviewed guide.Source for Limits & data boundaries (opens in a new tab)
Cache reads, writes & storage
ChargeApplies toRate / unit
Cached input1CachingR$0.25 / 1M tokensPricing evidence (opens in a new tab)

Rate conditions

  1. BRL per million tokens. Cache hits retain their own 75% input discount. Batch, Flex and night rates apply only to uncached input and output and are mutually exclusive. Night hours are 22:00–06:00 BRT. Flex is capacity-dependent and may queue up to five minutes or return 429. No currency conversion is applied.

Observed 14 Oct 2026. Write rates and storage fees are different charges; use the billing rules above.

Caching reviewed 14 Oct 2026

Maritaca API / São Paulo

sabiazinho-4-br-sp
Documented
Support evidence (opens in a new tab)
How it works
Automatic prompt-prefix cacheSource for How it works (opens in a new tab)
Minimum reusable input
Not collected

Minimum eligible prefix length is not specified.

Lifetime & refresh
Not collected

No retention/TTL or refresh policy is published in the reviewed cache guide.

How to enable it
No explicit cache creation; keep stable content first and reuse identical token prefixes.Source for How to enable it (opens in a new tab)
Usage fields
prompt_tokens_details.cached_tokensSource for Usage fields (opens in a new tab)
Billing rules
Cached input costs 25% of standard input; discounts on uncached input/output do not compound with cache.Source for Billing rules (opens in a new tab)
Limits & data boundaries
Batch, Flex and night pricing are mutually exclusive. Cache scope, cross-account isolation and regional reuse are not specified in the reviewed guide.Source for Limits & data boundaries (opens in a new tab)
Cache reads, writes & storage
ChargeApplies toRate / unit
Cached input1CachingR$0.33 / 1M tokensPricing evidence (opens in a new tab)

Rate conditions

  1. BRL per million tokens. Cache hits retain their own 75% input discount. Batch, Flex and night rates apply only to uncached input and output and are mutually exclusive. Night hours are 22:00–06:00 BRT. Flex is capacity-dependent and may queue up to five minutes or return 429. No currency conversion is applied. BR-SP token charges include the published 30% premium; cached-input figures are the rounded published values, not an inferred higher-precision rate.

Observed 14 Oct 2026. Write rates and storage fees are different charges; use the billing rules above.

Caching reviewed 14 Oct 2026

What the measurements say

OAB-Bench legal drafting

Maritaca AI · vendor-reported
7.02score / 10

January 2026 provider launch evaluation. Legal drafting uses 0–10 scoring; agent result combines pass³ and success@1 across four environments. Exams cover 13 tests; Multi-IF averages three turns. BRACEval is pairwise win rate, not the old 10-point grading metric. The March technical report rounds or revises some Sabiazinho scores; those are not silently substituted.

Evaluation source (opens in a new tab)

Magis-Bench judicial drafting

Maritaca AI · vendor-reported
4.5score / 10

January 2026 provider launch evaluation. Legal drafting uses 0–10 scoring; agent result combines pass³ and success@1 across four environments. Exams cover 13 tests; Multi-IF averages three turns. BRACEval is pairwise win rate, not the old 10-point grading metric. The March technical report rounds or revises some Sabiazinho scores; those are not silently substituted.

Evaluation source (opens in a new tab)

Brazilian legislation

Maritaca AI · vendor-reported
85% accuracy

January 2026 provider launch evaluation. Legal drafting uses 0–10 scoring; agent result combines pass³ and success@1 across four environments. Exams cover 13 tests; Multi-IF averages three turns. BRACEval is pairwise win rate, not the old 10-point grading metric. The March technical report rounds or revises some Sabiazinho scores; those are not silently substituted.

Evaluation source (opens in a new tab)

Agentic capabilities, four environments

Maritaca AI · vendor-reported
55.2aggregate score / 100

January 2026 provider launch evaluation. Legal drafting uses 0–10 scoring; agent result combines pass³ and success@1 across four environments. Exams cover 13 tests; Multi-IF averages three turns. BRACEval is pairwise win rate, not the old 10-point grading metric. The March technical report rounds or revises some Sabiazinho scores; those are not silently substituted.

Evaluation source (opens in a new tab)

Brazilian exams

Maritaca AI · vendor-reported
81.5% accuracy

January 2026 provider launch evaluation. Legal drafting uses 0–10 scoring; agent result combines pass³ and success@1 across four environments. Exams cover 13 tests; Multi-IF averages three turns. BRACEval is pairwise win rate, not the old 10-point grading metric. The March technical report rounds or revises some Sabiazinho scores; those are not silently substituted.

Evaluation source (opens in a new tab)

Portuguese Multi-IF

Maritaca AI · vendor-reported
81.4% strict instruction following

January 2026 provider launch evaluation. Legal drafting uses 0–10 scoring; agent result combines pass³ and success@1 across four environments. Exams cover 13 tests; Multi-IF averages three turns. BRACEval is pairwise win rate, not the old 10-point grading metric. The March technical report rounds or revises some Sabiazinho scores; those are not silently substituted.

Evaluation source (opens in a new tab)

BRACEval

Maritaca AI · vendor-reported
66.5% wins against GPT-4o

January 2026 provider launch evaluation. Legal drafting uses 0–10 scoring; agent result combines pass³ and success@1 across four environments. Exams cover 13 tests; Multi-IF averages three turns. BRACEval is pairwise win rate, not the old 10-point grading metric. The March technical report rounds or revises some Sabiazinho scores; those are not silently substituted.

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

Provider explicitly states Sabiá 4 models are API-only and offers neither weight downloads nor local/on-premise licensing.

Classification evidence (opens in a new tab)

Lifecycle dates

Public preview announced5 Jan 2026Source (opens in a new tab)
BR-SP endpoint opened to all API customers24 Jun 2026Source (opens in a new tab)

Sources reviewed