Four open-weight models spanning edge deployment to frontier capabilities. Every model ships with complete architecture documentation, training methodology papers, and permissive licensing for commercial use.
7 Billion Parameters
NAPH-7B is our most compact model, designed specifically for resource-constrained environments where latency and memory efficiency take priority. Despite its smaller footprint, it maintains strong performance on core language tasks through careful architecture optimization and high-quality training data curation.
The model runs comfortably on consumer-grade hardware including gaming laptops with 8GB VRAM when using 4-bit quantization. This makes it ideal for on-device applications, offline processing, and scenarios where internet connectivity is unavailable or undesirable.
70 Billion Parameters
NAPH-70B represents the optimal balance between capability and practical deployment. It delivers performance competitive with leading closed-source models while remaining deployable on standard enterprise GPU infrastructure. This is our most widely deployed model.
The extended 128K context window enables processing of entire codebases, lengthy legal documents, and complex multi-turn conversations without truncation. Native tool use support makes it suitable for agentic applications requiring structured function calling.
405 Billion Parameters
NAPH-405B is our frontier model, designed to compete with the most capable systems from any AI lab. It achieves state-of-the-art performance across reasoning benchmarks, complex analysis tasks, and creative generation while maintaining full transparency into its architecture and training.
The model excels at tasks requiring deep reasoning, nuanced understanding, and sophisticated output. It's particularly strong on mathematical reasoning, scientific analysis, and complex instruction following. Multi-modal capabilities support image understanding alongside text.
34 Billion Parameters
NAPH-Coder is purpose-built for software development tasks. Rather than training a general model and hoping it performs well on code, we designed the architecture, curated the training data, and optimized the training process specifically for programming applications.
The model understands 80+ programming languages with deep knowledge of idioms, patterns, and best practices. It excels at code completion, generation, review, refactoring, documentation, and debugging. Extended context allows processing entire codebases for architectural understanding.
Performance
Performance across standard evaluation benchmarks. All results are from internal evaluations using consistent methodology. We publish full evaluation details in our technical reports.
| Benchmark | NAPH-7B | NAPH-70B | NAPH-405B | NAPH-Coder |
|---|---|---|---|---|
| MMLU (5-shot) | 63.2% | 79.8% | 86.4% | 71.3% |
| HumanEval (pass@1) | 42.1% | 67.4% | 81.2% | 89.6% |
| GSM8K (CoT) | 58.7% | 86.3% | 94.1% | 78.9% |
| MATH (4-shot) | 24.6% | 48.2% | 61.8% | 43.5% |
| HellaSwag | 79.4% | 87.2% | 91.3% | 84.1% |
| MBPP (pass@1) | 48.3% | 72.1% | 84.7% | 91.2% |
| WinoGrande | 74.8% | 83.6% | 88.4% | 79.2% |
Get API access or talk to our team about on-premises deployment options.