MariTalk API

Sabiá is a family of Portuguese-specialized models developed by Maritaca AI, available via API. These models are designed to simulate high-fidelity human interactions in Portuguese, covering tasks such as question answering, context-aware information retrieval, translation, creative text generation, and data analysis.

Key Characteristics:

• Language Specialization: Optimized specifically for Portuguese.
• Billing Model: Usage is billed based on token volume (both input and output tokens).
• Capabilities: Includes text generation, translation, tutoring, customer support, and data interpretation.

Note: Sabiá models operate based on learned language patterns and do not possess consciousness or real understanding.

Maritaca's Batch API allows you to perform high-volume, asynchronous operations. It is divided into two main functional areas:

1. Files (/api/files): Used to store, retrieve, download, or delete files (such as .jsonl, .csv, or .txt) that serve as inputs for batch jobs.
2. Batch Inference (/api/batches): Used to run asynchronous jobs in bulk using previously uploaded files.

Base URL: https://chat.maritaca.ai

Sabiá 4 is optimized for Brazilian legal, educational, and institutional applications. It provides high accuracy on Brazilian legislation and improved agentic capabilities compared to previous versions (Sabiá 3.1).

Key Performance Areas:

• Brazilian Laws: High accuracy in knowledge of Brazilian legislation.
• Legal Tasks: Strong performance in OAB (lawyer) and Magis (judge) benchmarks.
• Agentic Capabilities: Improved tool use and environment interaction in Portuguese.
• Brazilian Exams: High accuracy across various national exams (ENEM, USP, etc.).
• Conversational Skills: Evaluated via BRACEval for Portuguese proficiency.

Sabiá 4 is positioned as a balanced choice for local context needs with a cost profile significantly lower than many global frontier models.

The Batch API is an asynchronous request mode designed for high-volume workloads that do not require immediate responses. It offers up to 50% cost reductions compared to standard requests.

Key Characteristics

• Cost & Limits: Billing occurs per token, but rate limits are calculated by characters per day.
  • Tiers 1-5: 1.2 billion characters/day (~300M tokens).
  • Tier 0: 4 million characters/day (~1M tokens).
  • Only input tokens count toward the rate limit.
• Latency: Requests can take up to 24 hours to complete.
• Expiration: Batches expire if not finished within the 24-hour window. Use with caution for critical scenarios.

Ideal Use Cases

• Translation or mass text review
• Large-scale evaluations
• Dataset classification
• Document summarization
• Entity and keyword extraction

MariTalk uses a tiered system where higher usage and spending automatically promote you to higher tiers with increased rate limits.

Important: Initial credits and coupons do not count towards tier upgrades; only actual API spending contributes to moving to the next tier.

Tier Requirements

• Tier 0: R$ 0 (To start testing)
• Tier 1: Any amount of paid usage
• Tier 2: R$ 100 spending
• Tier 3: R$ 500 spending
• Tier 4: R$ 2,000 spending
• Tier 5: R$ 5,000 spending

Performance varies based on the model version and available GPU hardware. Below are typical performance expectations:

Sabiá-2 Small

• On 1x A100 40GB: Expect ~118 tokens/s (total) for a 500 input / 100 output request (approx. 5 seconds total).
• On 2x A100 40GB: Higher throughput than single GPU setups.
• On 1x A10 24GB: Lower throughput, approximately 89.8 total tokens/s.
• Memory Requirement: ~20GB.

Sabiá-2 Medium

• On 2x A100 40GB: Expect ~79.3 total tokens/s. A request with 500 input / 100 output tokens takes approximately 10 seconds.
• Memory Requirement: ~70GB.

Note: As concurrency increases, total system throughput typically increases, but individual request latency (time to finish) also increases.

The 75% cache discount applies specifically to cached input tokens. Other discounts (Batch API, Flex, and off-peak) apply to non-cached input tokens and output tokens.

Note that Batch, Flex, and off-peak discounts are mutually exclusive; only one of them is applied per request. They do not multiply with each other, but they can coexist with the cache discount on different token types.

While the Chat Completions API remains available, the Responses API is the recommended interface for new projects. Key differences include:

MariTalk Local runs on 64-bit Linux machines with one or more NVIDIA GPUs (Compute Capability 8.0+). Memory requirements vary by model:

Notes:

• GPU memory can be distributed across multiple devices (e.g., for Medium, you can use one 80GB A100, two 40GB A100s, or four 24GB A10s).
• CPU memory is required for model processing before transfer to GPU; this memory is released after loading.

Sabiá 4 Thinking is the reasoning-specialized model within the Sabiá family. It is designed to achieve frontier-level quality specifically for Portuguese and Brazilian contexts.

Key characteristics include:

• High Performance: Significant gains in tool use (function calling), legal tasks, and overall answer quality compared to the standard Sabiá 4.
• Cost Efficiency: It is significantly more affordable than leading frontier models. Running a full suite of benchmarks on Sabiá 4 Thinking costs less than half of GPT-5.4 and approximately one-third of Opus 4.8.