BigBang-v1 lets AI generate its own harder training data

BigBang-v1 is an experimental 35B multimodal model that tests a fascinating premise: what if AI creates its own increasingly brutal curriculum, completely bypassing human-made tests?

Released by endless-frontier on Hugging Face on August 2, 2026, the model builds on top of Qwen 3.6 35B-A3B. By using an adversarial synthetic data pipeline to generate roughly 10,000 frontier scientific challenges, this 35B model ended up scoring between DeepSeek V4 Flash at 284B and DeepSeek V4 Pro at 1.6T on aggregate benchmarks.

A closed loop of generator and critic agents

As frontier language models hit the limits of human knowledge, hand-crafting post-training datasets has become a massive bottleneck. BigBang-v1 tackles this with an automated adversarial system.

A generator agent is tasked with continuously proposing and solving complex scientific and technical problems. Meanwhile, a critic agent audits every problem for correctness, genuine difficulty, diversity, and scalability, filtering out shallow examples. The pipeline also uses held-out research tasks to calibrate the generated distribution against real-world science.

Through this back-and-forth loop, the system gathered roughly 10,000 top-tier training examples. The resulting jump was massive. On Frontier Science Research, the base Qwen model scored 11.9, while BigBang-v1 jumped to 46.2. On Humanity's Last Exam, PaperBench, and BioMystery-HD, it even outperformed the 1.6T DeepSeek V4 Pro Preview.

Massive context and deployment specs

BigBang-v1 comes configured with an out-of-the-box context window of 262,144 tokens. The team advises keeping at least 128K active during inference, as shrinking the window any further degrades its extended reasoning process.

The model handles both text and vision inputs, supports tool-use parsing, and features multi-token prediction. If you only plan to run scientific text reasoning, you can strip the vision encoder in vLLM to free up KV cache memory.

Serving the model at its full 262K context typically demands an 8-GPU tensor parallel setup running vLLM or SGLang. Weights are public on Hugging Face under endless-frontier/BigBang-v1, ready to run via Transformers, SGLang, vLLM, or standard Docker model runners.

Why it matters

For AI researchers and builders tackling complex technical agents, BigBang-v1 is worth downloading right away. It proves that post-training quality trumps raw quantity: 10,000 self-generated, verifiable frontier tasks can push a 35B base model into heavyweight territory.

While calling it self-evolving is a bit of a stretch—the system is essentially an automated data flywheel that will eventually hit diminishing returns once the generator exhausts its baseline intuition—it still offers a compelling blueprint. Small labs no longer need massive crowdsourced labeling farms to train competitive reasoning models.