A naukri.com initiative
VentureBeat
2w
372
Image Credit: VentureBeat
Microsoft launches Phi-4-Reasoning-Plus, a small, powerful, open weights reasoning model!
- Microsoft Research has released Phi-4-reasoning-plus, a 14-billion parameter language model designed for tasks requiring deep reasoning.
- The model integrates supervised fine-tuning and reinforcement learning to enhance performance in math, science, and logic-based tasks.
- Phi-4-reasoning-plus is licensed under MIT, compatible with various frameworks, and aims for high-quality reasoning.
- Structured training with special tokens enhances transparency and coherence in problem solving.
- The model balances correctness, conciseness, and formatting consistency through reinforcement learning.
- Phi-4-reasoning-plus is suited for applications requiring high-quality reasoning under memory or latency constraints.
- It supports context lengths up to 64,000 tokens and performs well in chat-like settings.
- Despite its smaller size, the model outperforms larger models on benchmarks like the AIME 2025 math exam.
- Phi-4-reasoning-plus is positioned as a research tool for generative AI systems and not as a universal solution.
- Careful evaluation of performance, safety, and fairness is advised before deploying the model in high-stakes environments.
Read Full Article
22 Likes
For uninterrupted reading, download the app