35.95B total / approximately 3B active (MoE)
262,144 tokens
Apache License 2.0
General-purpose multimodal model evolved from Qwen3.6-35B-A3B through post-training for scientific research, reasoning, coding, and tool use. BigBang-V1 was trained on approximately 10,000 high-difficulty examples produced by a self-evolving synthetic-data framework in which generator agents create and solve problems while critic agents assess correctness, difficulty, scalability, and diversity. It uses a 40-layer hybrid architecture that repeats three linear-attention layers followed by one full-attention layer. Its Mixture-of-Experts design contains 256 experts, selects eight experts per token, and activates approximately 3B parameters. The model supports reasoning, structured tool calls, multimodal image and video input, and multi-token prediction. It includes a 27-layer vision encoder and can be deployed in text-only mode without loading the vision encoder. The Endless Frontier reports scores of 76.5 on BrowseComp, 54.2 on SWE-Bench Pro, 50.3 on Humanity's Last Exam, and 46.2 on FrontierScience Research. Official deployment guidance covers Transformers, SGLang, vLLM, and KTransformers.