Coding
The biggest leap of the family: superior C++/Rust synthesis, more accurate and coherent code generation, and better understanding of full-scale codebases.
Launched August 26, 2026
Sylor S2 Pro.
Orzatty's new flagship: 744B parameters in a MoE architecture with only 40B active per token, native 1M token context, and the family's biggest leaps in coding and agentic tasks.
New · September 2, 2026
The GGUFs are now on Hugging Face.
Sylor S2 Pro quantized weights are published in GGUF format for local execution with llama.cpp, LM Studio and Ollama. Download them today from Hugging Face.
Download from Hugging FaceSpecifications
Technical sheet.
Sylor S2 Pro is the second generation of the largest model built in Latin America. A next-generation Mixture-of-Experts architecture focused on efficiency and reasoning.
Total parameters 744B
Active parameters 40B
Architecture Mixture-of-Experts
Native context 1M tokens
Open format GGUF
Launch date August 26, 2026
The OrzattyCloud platform currently exposes a 256K limit.
Key improvements
Where it leaps.
The second generation is not just bigger: it is dramatically better at the two capabilities that matter most today for autonomous software development.
Agentic tasks
Long-horizon planning, tool use and more stable multi-step reasoning. Designed to power autonomous agents end to end.
1M context
A native one-million-token context window: entire codebases, complete documentation and long conversations in a single pass.
MoE efficiency
40B active parameters per token: frontier capacity with a radically lower inference cost than an equivalent dense model.
Comparison
S1 Pro vs S2 Pro.
A year of research between both generations. The leap shows.
| Sylor S1 Pro | Sylor S2 Pro | |
|---|---|---|
| Total parameters | 1T (MoE) | 744B (MoE) |
| Active parameters | 32B | 40B |
| Context | Standard | 1M native |
| Coding | Strong | Generational leap |
| Agentic | Basic | Built for agents |
| Open format | GGUF | GGUF |
Run Sylor S2 Pro today.
Download the GGUFs from Hugging Face or access via platform.