Chinese Team Debuts BigBang-V1: A 35B Native RSI Model Beating DeepSeek V4
新智元 · wechat · 2026-08-06
The 'Endless Frontier Team' comprising Shanghai Jiao Tong University and DP Technology, debuted BigBang-V1, the first 35B foundation model trained via native Recursive Self-Improvement (RSI). Outperforming DeepSeek V4 Pro (1.6T parameters) in multiple hard science benchmarks, it proves parameter scaling isn't the only path forward.
The RSI Loop
BigBang-V1's breakthrough lies in its 100% AI-synthesized data pipeline. It uses Generator and Critic agents wrapped in a real-world testing loop. If the Critic rates data highly but the model fails to improve, the system uses actual results to recalibrate the Critic, creating a closed loop of continuous self-improvement.
Robust Scientific Reasoning
- Benchmarks: Scored 46.2 on OpenAI's Frontier Science Research and 50.3 on Humanity's Last Exam.
- Practical Skills: Can accurately identify viruses from raw data, reproduce academic papers (even fixing logical flaws in the original text), and orchestrate open-source models to design antibodies.
The team emphasizes that scientific tasks, due to their frontier nature and verifiability, are the ultimate training ground for AGI.
More from Models
- DeepSeek Announces Significant API Price Hikes Amid Surging Demand — rohanpaul_ai · 2026-08-06
- Claude Opus 5 Takes #1 Spot in Fullstack Code Arena, Beating Kimi K3 Max — rohanpaul_ai · 2026-08-06
- API Data Shows OpenAI Models Beat Claude in Cost-Efficiency Across All Tiers — Over-Necessary-4774 · 2026-08-06
- Factoring in Retries: Are Cheaper AI Models Actually Cheaper? — ExtremeAdmirable4097 · 2026-08-06
- Nebius Inference Platform Hits Milestone in Artificial Analysis Accuracy Index — demian_ai · 2026-08-06
- First Fine-tunes of LFM2.5 Released: Macaw On-Device Agent and BTL-4 Reasoning Model — maximelabonne · 2026-08-06