Show HN: 8B Parallel Coordinated Reasoning Model (github.com)

The researchers have unveiled the PaCoRe (Parallel Coordinated Reasoning) framework, a transformative approach to AI inference that prioritizes coordinated parallel reasoning over traditional sequential depth. This innovative model allows for expansive parallel exploration during inference, effectively increasing the model's reasoning capabilities by overcoming context limitations. Notably, the PaCoRe-8B model achieved a remarkable score of 94.5% on the HMMT 2025 benchmark, surpassing GPT-5's 93.2%, signaling significant advancements in AI reasoning accuracy across various domains. PaCoRe employs a message-passing architecture, synthesizing insights from multiple parallel trajectories into concise messages that inform subsequent reasoning rounds. Trained through large-scale reinforcement learning, this model showcases the potential of parallel computation, unlocking superior gains in reasoning efficiency as test-time compute increases. The project is open-source, including model checkpoints and training data, which invites further research and development in parallel reasoning techniques. By emphasizing breadth over depth and providing robust data resources, PaCoRe is poised to accelerate future advancements in AI, including the exploration of emergent multi-agent intelligence and enhanced training methodologies.