SJTU's BigBang-v1 Beats a 1.6T Model Using Only Machine-Generated Training Data
A 35B open model post-trained on self-generated adversarial tasks matches trillion-parameter systems on frontier science and coding benchmarks, now packaged as GGUF for local inference.
Takeaways
- Endless Frontier released BigBang-v1, a 35B MoE built on Qwen 3.6 35B-A3B with a 262K context window.
- Post-trained on fully AI-synthesized, adversarial, self-evolving verifiable tasks instead of scraped data.
- Claims top scores among 35B models on 8-10 benchmarks; beats DeepSeek V4 Pro (1.6T) on FrontierScience, HLE, PaperBench.
- Bartowski published GGUF imatrix quants from BF16 (71GB) down to IQ2_XXS (10GB); Q4_K_M recommended at 21.86GB.
- Includes multimodal mmproj for image input and MTP layers for speculative decoding via
--spec-type draft-mtp. - Apache 2.0 licensed; runs in llama.cpp b10262+, LM Studio, Ollama, vLLM, and SGLang 0.5.10+.
BigBang-v1, a 35B model from the Endless Frontier team at Shanghai Jiao Tong University's AI Institute, landed on Hugging Face claiming scores at or above trillion-parameter frontier models on scientific research, coding, and long-horizon agent benchmarks. Bartowski followed with a full set of GGUF quantizations runnable on a single workstation, and the community has pulled it hundreds of thousands of times. The model's training method is what makes the benchmark numbers worth taking seriously: every training task was generated, filtered, and iterated by AI, with no human-written data in the loop.
35B beating 1.6T on frontier tasks
BigBang-v1 is a 35.1 billion parameter mixture-of-experts model, post-trained from Qwen3.6-35B-A3B. According to the team's GitHub README, it achieves the highest reported score among 35B-scale models across eight benchmarks and exceeds DeepSeek V4 Pro Preview (1.6T) on FrontierScience Research, Humanity's Last Exam, PaperBench (Code-Dev), and BioMysteryBench-HD. The evaluation suite spans long-horizon search, software engineering, scientific research, and AI research tasks. Default context length is 262,144 tokens.