The next round of large-model competition may begin in the task set, not in the parameter table.
On August 9, QbitAI reported the release of BigBang-V1 from The Endless Frontier team. The model is post-trained from Qwen3.6-35B-A3B, has 35B total parameters, activates about 3B during inference, and carries a default 262,144-token context window. The model, code, and technical report are public.
The task generator becomes part of training
The model card describes an adversarial, self-evolving synthetic-data framework. Generator agents propose and solve increasingly difficult scientific and technical problems. Critic agents check correctness, difficulty, scalability, and diversity, with held-out real research tasks used to calibrate the data distribution.
The reported scale is about 10,000 high-difficulty post-training examples. The point is not bulk generation. The data pipeline changes fields, chain length, tool use, and filtering strategy as the model improves.
“AI advancing science, and science advancing AI.”
That slogan only works if the tasks are verifiable. BigBang leans on problems that can be checked by code execution, numerical computation, simulation, formal methods, or domain tools.
Read the scores by protocol
The headline numbers are strong. BigBang-V1 is reported as the top selected 35B model across eight representative benchmarks, and it exceeds DeepSeek V4 Pro Preview on FrontierScience Research, Humanity's Last Exam, PaperBench(Code-Dev), and BioMysteryBench-HD.
Specific checks include 76.5 on BrowseComp, 54.2 on SWE-Bench Pro, 46.2 on FrontierScience Research, and 53.6 on PaperBench(Code-Dev). Those figures are also why the comparison needs care: BrowseComp, xbench, FrontierScience Research, and HLE use the repository's general-agent runner, while SWE-Bench Pro and SciCode-Verified use benchmark-specific official harnesses.
A safer reading is that data-layer self-evolving tasks can push a 35B-A3B base model materially upward on long-horizon search, coding, and research work.
The verification chain matters most
The repository documents an evaluation-time decontamination policy: Hugging Face pages are filtered from search, direct visits to those domains are blocked, and reference answers are not exposed to the agent. The general-agent loop uses search, visit, and code execution, with each trajectory capped at 500 tool calls.
That does not settle every benchmark dispute, but it puts the harness and policy in view. The next test is third-party reproduction: running the open model, checking the long-context setup, and seeing whether the same data-production idea transfers beyond research, coding, and search.
For China's model ecosystem, this is a useful signal. Beyond parameter counts and price cuts, data production is becoming core infrastructure. If the BigBang route reproduces, future model progress may depend less on static datasets and more on verifiable task factories.
Sources: QbitAI, Hugging Face model card, CocoLoop, GitHub repository, Endless Frontier technical report; verification covers BigBang-V1 parameter scale, 3B active parameters, 262K context, about 10,000 post-training examples, benchmark scores, evaluation harnesses, and decontamination policy.