Osakra/Project-Norn-V17-4.5B
Osakra/Project-Norn-V17-4.5B is a 4.45 billion parameter neuro-symbolic dual-system intelligence model developed by Osakra Research. It combines a 16-bit causal language model with a 1024-dimensional 4-channel Relational Graph Holographic Reduced Representation (HRR) engine for continuous latent deliberation. This model excels in complex reasoning tasks, particularly algorithmic code generation, achieving a 92.0% polyglot code compilation rate in its cognitive mode, and supports a 32K context length.
Loading preview...
Overview
Project Norn V17 is a 4.45 billion parameter neuro-symbolic dual-system intelligence model from Osakra Research. It integrates a 16-bit causal language model with a unique 1024-dimensional 4-channel Relational Graph Holographic Reduced Representation (HRR) engine. This architecture allows for two operational modes: a fast, standalone causal LM (Mode A) and a deep, cognitive engine (Mode B) that performs continuous latent deliberation before token generation. Unlike previous versions, V17 includes pre-fused 16-bit base weights, enabling out-of-the-box compatibility with standard Hugging Face tools, vLLM, llama.cpp, and Ollama.
Key Capabilities
- Dual-System Reasoning: Offers both a high-speed standalone causal LM and a more deliberate neuro-symbolic cognitive engine for complex tasks.
- Enhanced Code Generation: Mode B (Cognitive Engine) significantly boosts polyglot code synthesis, achieving a +10.0% uplift to 92.0% AST parse rate by using relational graph compilation.
- Relational Graph Memory: Utilizes a 1024-dimensional HRR engine with four specialized channels for topological, algebraic, causal, and goal-attractor encoding, enabling continuous latent Chain-of-Thought.
- Adaptive Computation Time (ACT): Features an epistemic halting head that dynamically adjusts deliberation depth, preventing over-computation on simpler queries.
- High Benchmark Performance: Achieves 76.40% Micro Accuracy in standalone mode and 75.60% in cognitive mode on a 250-sample held-out benchmark, with strong results across GSM8K Math, MMLU, and Agentic Sandbox tasks.
- Conversational Quality: Resolves conversational degradation, demonstrating 0.0% code-fence bias on dialogue and 100.0% self-knowledge accuracy.
- Reproducible Evaluation: Bundles a complete 250-sample evaluation suite and scripts for full scientific reproducibility.
Good For
- Developers requiring advanced algorithmic code generation and complex problem-solving, especially where relational understanding is critical.
- Applications needing a compact yet powerful model capable of deep reasoning without excessive token latency.
- Use cases benefiting from flexible deployment options, including direct Hugging Face integration, GGUF conversion, and local API serving (OpenAI/Ollama protocols).