This technical report presents Sabiá-4 and Sabiazinho-4, a new generation of Portuguese language models with a focus on Brazilian Portuguese language. The models were developed through a four-stage training pipeline: continued pre-training on Portuguese and Brazilian legal corpora, long-context extension to 128K tokens, supervised fine-tuning on instruction data spanning chat, code, legal tasks, and function calling, and preference alignment. We evaluate the models on six benchmark categories: conversational capabilities in Brazilian Portuguese, knowledge of Brazilian legislation, long-context understanding, instruction following, standardized exams, and agentic capabilities including tool use and web navigation. Results show that Sabiá-4 and Sabiazinho-4 achieve a favorable cost-performance trade-off compared to other models, positioning them in the upper-left region of the pricing-accuracy chart. The models show improvements over previous generations in legal document drafting, multi-turn dialogue quality, and agentic task completion.
Figures & tables
Figure 1: Total inference cost versus average benchmark accuracy. Optimal models appear in the upper-left quadrant (cheaper and better). Sabiá models highlighted in green. Currency conversion: BRL/USD = 5.4.
Figure 2: Training pipeline overview. Pre-training consists of Portuguese continual learning on general and legal corpora, followed by long context training. Post-training includes supervised fine-tuning on diverse instruction data and preference alignment.
Benchmark
Description
Metric
OAB Bench
Legal Drafting (Attorney Style)
Average score [0, 10]
Magis Bench
Legal Drafting (Judge Style)
Average score [0, 10]
Brazilian laws
Knowledge of Brazilian law
Accuracy (5 alternatives)
Agentic capabilities
Tool usage in four environments in Portuguese
Pass^3 and success@1
Brazilian exams
13 exams (ENEM, CFC, Revalida, CPNU, OAB etc)
Accuracy (4 and 5 alternatives)
Portuguese Multi-IF
Instruction-following capability
Strict Average over 3 turns
Table 1: List of benchmarks used to evaluate model capabilities across different domains and tasks.
Benchmark
sabiazinho-4
gpt-oss-120b
gpt-4.1-mini
gpt-5-mini
gemini-2.5-flash-lite
OAB-Bench (Lawyer Evaluation)
7.02
6.01
5.50
6.37
6.25
Magis-Bench (Judge Evaluation)
4.50
3.62
3.67
4.47
4.25
Laws (Legal Knowledge)
85.0
52.3
57.0
68.2
72.1
Agent Capabilities (4 environments)
55.2
60.9
59.4
85.1
18.0
Multiple Choice Exams (13 exams)
81.0
77.0
81.0
84.6
76.2
Multi-IF PT (Instruction Following)
81.0
82.0
79.6
85.8
80.8
Table 2: Comparison of models for the same price range: cost-effective models.
Benchmark
sabia-3.1
sabia-4
Qwen3 235b
gpt-4.1
gpt-5.2 (instant)
gpt-5.2 (high)
Gemini-3 Pro (low)
Gemini-3 Pro (high)
kimi-k2 thinking
deepseek v3.2
OAB-Bench
7.21
7.49
6.33
7.30
8.07
8.73
9.05
8.90
6.62
6.40
Magis-Bench
4.97
5.08
4.52
5.55
6.66
6.99
7.79
7.48
4.49
4.88
Laws
77.8
97.4
65.9
80.8
84.0
86.3
74.9
88.6
59.1
67.3
Agent Capabilities
43.1
72.2
67.8
73.3
81.1
85.7
90.4
90.1
77.3
40.5
Multiple Choice Exams
82.4
86.6
82.0
86.1
88.0
92.9
93.3
95.0
83.0
84.0
Multi-IF PT
80.7
82.0
84.4
82.7
83.7
87.2
86.0
88.0
86.0
81.5
Table 3: Comparison of models for the same price range: frontier models
Figure 3: Comparison of Sabiá generations on the MRCR benchmark. Sabiá-4 and Sabiazinho-4 are the new model generations.
Figure 4: Sample from BRACEVal benchmark. Note that the example was translated to English for the paper. Original language is Portuguese.
Figure 5: Sample from Climb benchmark. Note that the example was translated to English for the paper. Original language is Portuguese.
Figure 6: Sample from Magis benchmark. Note that the example was translated to English for the paper. Original language is Portuguese.
Figure 7: Sample from OAB benchmark. Note that the example was translated to English for the paper. Original language is Portuguese.
Figure 8: Sample from Ticket-bench benchmark. Note that the example was translated to English for the paper. Original language is Portuguese.
Figure 9: Sample from Brazilian laws benchmark. Note that the example was translated to English for the paper. Original language is Portuguese.
Figure 10: Sample from MARCA benchmark. Note that the example was translated to English for the paper. Original language is Portuguese.
Figure 11: Sample from Multi-IF benchmark. Note that the example was translated to English for the paper. Original language is Portuguese.
Figure 12: Sample from pix-bench benchmark. Note that the example was translated to English for the paper. Original language is Portuguese.
Model
Input [$/1m]
Output [$/1m]
Provider
Sabiazinho-4
0.19
0.74
Maritaca AI
gpt-oss-120b
0.15
0.6
Together AI
gpt-4.1-mini
0.4
1.6
Openai
gemini-2.5-flash-lite
0.1
0.4
Google
gpt-5-mini
0.25
2
OpenAI
sabiá-3.1
0.93
1.85
Maritaca AI
Table 4: Input and output prices used for evaluations, including the providers considered.