Open-Source Model Wins First Physics Olympiad Gold: Shanghai AI Lab's 235B Model Beats GPT-5 and Grok-4

🏅 Open-source AI Model Wins Gold at International Physics Olympiad

Historic Achievement

The P1-235B-A22B model from Shanghai AI Lab has achieved a 21.2/30 score at the International Physics Olympiad (IPhO) — surpassing the gold medal threshold and making history as the first open-source model to win gold.

In the HiPhO benchmark (13 top-tier global physics competitions, 2024–2025), P1-235B-A22B earned:

• 12 gold medals
• 1 silver medal
• Tied first place on the leaderboard with Google Gemini-2.5-Pro

This surpasses:

• GPT-5 – 11 golds
• Grok-4 – 10 golds

It demonstrates that open-source models have now matched — and even surpassed closed-source models — in physics reasoning.

---

🌍 Significance of AI in Physics Reasoning

Physics reasoning is critical for understanding and shaping the real world. Prestigious competitions such as IPhO require:

• Complex reasoning
• Deep physics understanding

Winning gold is a vital milestone toward general physics intelligence and showcases real-world problem-solving potential.

---

🧪 HiPhO: Benchmark for Physics Olympiads

HiPhO (High School Physics Olympiad) is the first benchmark dedicated to recent Olympiad-level physics contests with human-aligned evaluation.

Coverage (2024–2025):

• IPhO
• APhO
• EuPhO
• Other regional Olympiads (total: 13 contests)

Evaluation Approach:

• Official competition scoring standards
• Fine-grained human-aligned evaluation of both answers and reasoning steps
• Scores directly comparable to human contestants’ medal boundaries

HiPhO benchmark overview (13 competitions worldwide).

Training Dataset:

• Thousands of Olympiad-level problems
• Full context + verifiable answers
• Standard solution paths
• Purpose-built for reinforcement learning training

---

📈 Multi-stage Reinforcement Learning in P1

P1-series models achieve sustained improvement via multi-stage reinforcement learning, with two core strategies:

• Expanding Context Window
• Gradual increase of output length
• Enables longer chains of reasoning
• Improves complex problem-solving & reduces truncation errors
• Pass-rate Filtering
• Uses pass-rate statistics before training
• Filters out tasks that are too easy or too hard

---

🤝 PhysicsMinions: Evolutionary Multi-agent Reasoning

To overcome single-model limitations, the team built PhysicsMinions — a collaborative evolutionary multi-agent system. It consists of three interconnected modules:

• Visual Module (Visual Studio)
• Observes and verifies multimodal problems
• Extracts structured visual information
• (Not used in P1 model’s experiments)
• Logic Module (Logic Studio)
• Generates initial solutions
• Iteratively revises answers via self-reflection
• Review Module (Review Studio)
• Physics Validator: checks physical consistency (constants, units)
• General Validator: checks logical and calculation soundness

❗ If a stage fails, an error report is sent back to the Logic Module for refinement — iterating toward higher accuracy.

PhysicsMinions framework and module interactions.

---

📊 Results & Benchmark Highlights

P1-235B-A22B:

• 12 gold + 1 silver (HiPhO)
• Surpassed: GPT-5 (11 gold), Grok-4 (10 gold)
• Scored 21.2/30 at IPhO 2025 — only open-source gold winner

P1-30B-A3B:

• 8 gold, 4 silver, 1 bronze (HiPhO)
• Ranked 3rd among open-source models
• Beats several closed-source models (e.g., o4-mini, Claude-4-Sonnet)

PhysicsMinions Impact:

• P1-235B-A22B avg score: 35.9 (HiPhO)
• With PhysicsMinions: 38.4 — overall 1st place surpassing Gemini-2.5-Pro (37.7) and GPT-5 (37.4)

P1 models also improved in math, code, STEM — showing generalization beyond physics.

---

🔗 Project Links

P1 Models:

• Project Page: https://prime-rl.github.io/P1
• GitHub: https://github.com/PRIME-RL/P1

HiPhO Benchmark:

• Paper: https://arxiv.org/abs/2509.07894
• Dataset: https://huggingface.co/datasets/SciYu/HiPhO
• Leaderboard: https://phyarena.github.io/

PhysicsMinions:

• Paper: https://arxiv.org/abs/2509.24855

---

🚀 Real-world Application for Creators

Platforms like AiToEarn官网 allow creators to:

• Generate AI-assisted content
• Publish across multiple social channels (Douyin, Kwai, WeChat, Bilibili, Xiaohongshu, Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, X/Twitter)
• Track analytics & performance
• Monetize creativity efficiently

By integrating outputs from advanced AI models such as P1 into content workflows, creators can achieve global reach and fast monetization — powered by open-source AI breakthroughs.