endless-frontier/BigBang-v1
endless-frontier/BigBang-v1 is a 35.1 billion parameter general-purpose LLM, evolved from Qwen 3.6 35B-A3B. It was post-trained using an adversarial, self-evolving synthetic data framework to generate verifiable frontier tasks. This model excels in scientific research, reasoning, coding, and tool-use benchmarks, achieving performance comparable to much larger models like DeepSeek V4 Flash (284B) and DeepSeek V4 Pro (1.6T). It is particularly optimized for solving increasingly challenging scientific and technical problems with a default context length of 262,144 tokens.
Loading preview...
BigBang-v1: Advancing Frontier AI through Self-Evolving Synthesis
BigBang-v1 is a 35.1 billion parameter large language model developed by endless-frontier, building upon the Qwen 3.6 35B-A3B architecture. Its core innovation lies in its training methodology: an adversarial, self-evolving synthetic data framework. This framework utilizes generator agents to propose and solve complex scientific and technical problems, while critic agents evaluate correctness, difficulty, and diversity, calibrating the synthetic data against real research tasks.
Key Capabilities & Differentiators
- Verifiable Frontier Tasks: BigBang-v1 is specifically designed to tackle problems at the boundary of current knowledge, where solutions can be objectively evaluated through formal methods, computation, or simulation.
- Superior Performance: Despite its 35.1B parameter count, BigBang-v1 demonstrates aggregate performance between DeepSeek V4 Flash (284B) and DeepSeek V4 Pro (1.6T) across scientific research, reasoning, coding, and tool-use benchmarks.
- Benchmark Leadership: It achieves the highest reported scores among selected 35B models on eight representative benchmarks, including long-horizon search, software engineering, scientific research (e.g., FrontierScience Research, Humanity's Last Exam), and AI research.
- Extended Context Window: The model supports a default context length of 262,144 tokens, crucial for complex tasks requiring extensive information processing.
Ideal Use Cases
BigBang-v1 is particularly well-suited for applications requiring advanced capabilities in:
- Scientific Research: Solving complex problems in various scientific domains.
- Software Engineering: Excelling in coding tasks and long-horizon search.
- Advanced Reasoning: Handling intricate logical and analytical challenges.
- Tool Use: Integrating and utilizing external tools effectively for problem-solving.