# BigBang-v1

> Source: https://aiwiki.ai/wiki/bigbang_v1
> Summary: BigBang-V1 is an open-weights large language model released on August 2, 2026 by The Endless Frontier, a Shanghai-based research team drawn from Shanghai Jiao Tong University's School of Artificial Intelligence, the AI-for-science company DP Technology, and the Institute for Advanced Algorithms…
> Updated: 2026-08-07
> Fact-checked: 2026-08-07
> Categories: AI Models, Chinese AI, Large Language Models, Open Source AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "BigBang-v1." aiwiki.ai, 7 Aug 2026. https://aiwiki.ai/wiki/bigbang_v1
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

| Field | Value |
|---|---|
| Developer | The Endless Frontier team |
| Contributing institutions | Shanghai Jiao Tong University School of Artificial Intelligence; DP Technology; Institute for Advanced Algorithms Research, Shanghai |
| Released | August 2, 2026 (Hugging Face weights) |
| Type | Mixture-of-experts large language model (agentic, multimodal input) |
| Parameters | 35B total, about 3B active per token |
| Base model | Qwen3.6-35B-A3B |
| Context length | 262,144 tokens |
| Training data (post-training) | ~10,000 AI-synthesized examples |
| License | Apache 2.0 |
| Website | endlessfrontier.tech |

**BigBang-V1** is an [open-weights](https://aiwiki.ai/wiki/open_weights) [large language model](https://aiwiki.ai/wiki/large_language_model) released on August 2, 2026 by The Endless Frontier, a Shanghai-based research team drawn from Shanghai Jiao Tong University's School of Artificial Intelligence, the AI-for-science company DP Technology, and the Institute for Advanced Algorithms Research, Shanghai [1][2][7]. The model is a post-trained derivative of Alibaba's [Qwen](https://aiwiki.ai/wiki/qwen)3.6-35B-A3B, a [mixture-of-experts](https://aiwiki.ai/wiki/mixture_of_experts) model with 35 billion total parameters and about 3 billion activated per token, and its distinguishing feature is its training data: roughly 10,000 post-training examples produced entirely by an adversarial, self-evolving [synthetic data](https://aiwiki.ai/wiki/synthetic_data) pipeline in which generator agents invent and solve "verifiable frontier tasks" and critic agents filter them [1][2]. The team reports that this comparatively small intervention lifts the 35B model into the aggregate performance range between DeepSeek V4 Flash Preview (284B parameters) and DeepSeek V4 Pro Preview (1.6T parameters) on a suite of science, coding, search, and AI-research benchmarks [2].

## Background

The project positions itself as a response to a widely discussed constraint on model progress: as models approach human-expert performance, training tasks conceived within the limits of existing human knowledge stop providing useful learning signal, whatever the raw volume of remaining internet text [2]. The report argues that continued open-ended capability growth requires "verifiable frontier tasks", meaning problems at the boundary of current knowledge whose candidate solutions can nonetheless be checked objectively through formal methods, computation, simulation, or domain-specific tools [2]. Scientific research is chosen as the task substrate because it continually generates new problems while offering many verification channels (proof checkers, program execution, numerical computation, simulators, databases, experimental feedback), and because solving a full scientific problem exercises search, evidence evaluation, hypothesis generation, coding, [tool use](https://aiwiki.ai/wiki/tool_use), and error recovery in combination [2].

The report frames its pipeline as an early, data-level form of recursive self-improvement: rather than a model improving its own weights or code, the system that produces training data improves its own task-generation and evaluation strategies over successive rounds [2]. It situates the work alongside prior AI-generates-its-own-training-signal efforts such as [Absolute Zero](https://aiwiki.ai/wiki/absolute_zero) and Google DeepMind's [AlphaEvolve](https://aiwiki.ai/wiki/alphaevolve), while claiming to differ by putting the task-generation program, task distribution, and evaluation criteria jointly under optimization, calibrated by training outcomes on real research tasks [2][7].

## The Endless Frontier team

The developer publishes under the name "The Endless Frontier" (the Hugging Face organization, created in 2026, listed 13 members and one model as of August 7, 2026) [10]. Chinese technology outlet Synced (机器之心) described the group as a team of researchers from Shanghai Jiao Tong University's School of Artificial Intelligence, DP Technology (深势科技), and the Institute for Advanced Algorithms Research, Shanghai (上海算法创新研究院) [7][12][13].

The technical report is credited on its title page simply to "The BigBang Team", with no individual author list; an appendix lists contributors by project area and names Siheng Chen and Linfeng Zhang, both with Shanghai Jiao Tong University email addresses, as project co-leads [2]. Siheng Chen is a tenure-track associate professor at Shanghai Jiao Tong University and a co-PI at the [Shanghai AI Laboratory](https://aiwiki.ai/wiki/shanghai_ai_lab) [14]; members of his group previously built the OpenSeeker open-source search-agent models, which the BigBang report cites as related work [2][15]. The appendix also lists five "Scientific Directors and Advisors", among them Weinan E, the applied mathematician known for pioneering machine learning applications in scientific computing [2][17].

## Model

BigBang-V1 keeps the architecture of its base model, Qwen3.6-35B-A3B (architecture class Qwen3_5MoeForConditionalGeneration), which Alibaba released under Apache 2.0 in April 2026 [5][9]. Like the base model, it accepts image and video inputs in addition to text (its Hugging Face pipeline tag is image-text-to-text), although the report's evaluations are text-based agentic tasks and the deployment guide includes a text-only serving mode that skips the vision encoder [1][5]. The weights are distributed in BF16 safetensors totaling 35.95 billion parameters (about 72 GB) under an Apache 2.0 license, with the license file inherited from the Qwen base model [1][5][6].

| Specification | Detail |
|---|---|
| Total parameters | 35.95B (BF16 safetensors) [5] |
| Active parameters | ~3B per token (A3B mixture-of-experts) [4] |
| Default context length | 262,144 tokens [1] |
| Modalities | Text, image, and video input; text output [5] |
| Recommended serving | SGLang >= 0.5.10, vLLM >= 0.19.0, KTransformers, or transformers serve [1] |
| Extras | Multi-token prediction (speculative decoding) configurations; Qwen3-style reasoning parser and tool-call parser [1] |

The model card recommends keeping at least 128K tokens of context "to preserve thinking capabilities", reflecting the model's design around [long-context](https://aiwiki.ai/wiki/long_context) agentic trajectories [1]. The GitHub repository additionally releases a lightweight general-agent harness exposing search, visit, and code_exec tools, along with the team's audited versions of evaluation datasets [4].

## Self-evolving training pipeline

BigBang-V1 was produced by post-training ([supervised fine-tuning](https://aiwiki.ai/wiki/fine_tuning), per the report's overview figure) on data from a two-level system the team calls a "training flywheel" [2].

The inner loop pairs two kinds of agents. A generator agent operates as a code agent that directly modifies, executes, and debugs the data-synthesis program itself, searching over task-construction, tool-use, verification, and sample-filtering strategies rather than merely sampling problems from prompts; after each round it records its motivations, results, and failure causes so later rounds can reuse the experience [2]. A critic agent adversarially reviews candidate data at two levels: basic constraint checks (complete reasoning traces, valid tool calls, parseable formats) and higher-level quality judgments (whether conclusions are supported by evidence, whether steps contradict each other, whether the final answer is objectively verifiable, and whether the task requires non-trivial research) [2].

Because critics can mistake tasks that merely look difficult for tasks that actually teach the model something, an outer loop calibrates the system against reality: model variants are trained on data from candidate pipeline versions and evaluated on held-out real research tasks, and a meta-critic compares the critic's quality predictions against the observed training outcomes, adjusting both the evaluation criteria and the generation strategy for the next round [2]. The full post-training set produced this way comprises approximately 10,000 high-difficulty examples spanning multiple scientific and technical domains, a figure stated in the model card and repeated in Chinese press coverage [1][7].

## Reported performance

All published numbers are the developers' own evaluations; as of August 7, 2026, no listing for BigBang-V1 had been identified on independent leaderboards such as LMArena or Artificial Analysis. The report evaluates the model as an agent with a 256K context window and up to 500 tool calls per trajectory, with access to Hugging Face domains blocked during browsing evaluations to prevent answer leakage [2].

Within the report's own comparison table, BigBang-V1 posts the highest score among the five compared 35B-scale models (including its Qwen3.6-35B-A3B base, Nex-N2 mini, Agents-A1, and Apodex-1.0-mini) on nine of eleven benchmark rows, ties Agents-A1 on a tenth (SciCode-V-Main), and trails Apodex-1.0-mini only on xbench [2]. It improves on its base model on all eleven rows, with the largest jumps on FrontierScience-Research (11.9 to 46.2) and xbench (32.6 to 58.4) [2]. Selected developer-reported scores [2]:

| Benchmark | Qwen3.6-35B-A3B (base) | BigBang-V1 | DeepSeek V4 Flash Preview | DeepSeek V4 Pro Preview | GPT-5.5 |
|---|---|---|---|---|---|
| [BrowseComp](https://aiwiki.ai/wiki/browsecomp) | 67.9 | 76.5 | 73.2 | 83.4 | 84.4 |
| xbench | 32.6 | 58.4 | 62.2 | 64.8 | 72.4 |
| [SWE-Bench](https://aiwiki.ai/wiki/swe_bench) Pro | 43.6 | 54.2 | 52.6 | 55.4 | 58.6 |
| [SciCode](https://aiwiki.ai/wiki/scicode)-V-Sub | 56.5 | 68.6 | 83.7 | 90.2 | 95.1 |
| SciCode-V-Main | 26.6 | 50.0 | 68.6 | 78.1 | 90.6 |
| FrontierScience-Research | 11.9 | 46.2 | 37.7 | 40.7 | 58.3 |
| [Humanity's Last Exam](https://aiwiki.ai/wiki/humanitys_last_exam) | 36.2 | 50.3 | 45.1 | 48.2 | 52.2 |
| BioMysteryBench (Human-Solvable) | 44.8 | 57.5 | 68.0 | 64.4 | 76.7 |
| BioMysteryBench (Human-Difficult) | 2.0 | 15.7 | 23.5 | 13.7 | 23.5 |
| [MLE-Bench](https://aiwiki.ai/wiki/mle_bench) (Lite) | 31.8 | 59.1 | 40.9 | 59.1 | 59.1 |
| [PaperBench](https://aiwiki.ai/wiki/paperbench) (Code-Dev) | 30.7 | 53.6 | 40.4 | 50.4 | 64.2 |

Against the larger [DeepSeek](https://aiwiki.ai/wiki/deepseek) V4 preview models, the team reports that BigBang-V1 beats the 284B-parameter Flash Preview on six of the eleven benchmarks and beats the 1.6T-parameter Pro Preview on four (FrontierScience-Research, Humanity's Last Exam, BioMysteryBench Human-Difficult, and PaperBench Code-Dev) [2]. The report itself notes that clear gaps to the strongest frontier systems remain on SciCode-Verified, BioMysteryBench, and the two AI-research benchmarks [2].

A separate inference-time "Search-and-Verify" procedure, which fans out one canonical and two complementary search trajectories and reconciles their compressed summaries, improves every reported benchmark it was applied to, most notably xbench (58.4 to 64.6) and BioMysteryBench Human-Solvable (57.5 to 63.0) [2].

The report's qualitative case studies emphasize verification behavior rather than raw answers: on a BioMysteryBench viral-identification task BigBang-V1 converged on the rubric-accepted Norovirus GII.4 answer in all three runs while DeepSeek V4 Flash produced three different incorrect virus labels, and on a symbolic elliptic-integral problem BigBang-V1 produced a fully derived closed-form answer in Gamma-function terms where DeepSeek V4 Pro Preview verified the same value numerically without completing the analytic derivation [2].

## Release and reception

The Hugging Face weights repository was created on August 2, 2026 and last updated on August 6; the GitHub repository followed on August 3 [4][5]. The paper is distributed as a self-published technical report on the project site (the site's own change notes describe the current version as v2, compiled August 2, 2026), while the GitHub page carried a "Paper Coming Soon" badge, and no arXiv listing had been identified as of August 7, 2026 [2][3][4].

Synced covered the release on August 7, 2026, framing it against Ilya Sutskever's argument that pre-training data is running out [7][8]. Adina Yakup of Hugging Face highlighted the model on X the same day [11]. Community quantizations followed, with GGUF conversions appearing within a day of release and an 8-bit MLX port by August 7 [16]. As of August 7, 2026 the official repository showed modest direct traction (on the order of 100 downloads and a few dozen likes), with the GGUF conversions accounting for roughly 2,000 additional downloads [5][16].

## Limitations

Every published benchmark figure is self-reported by the development team, and several of the comparison benchmarks (FrontierScience-Research, BioMysteryBench, the audited SciCode-Verified variant) are recent or team-audited, which complicates outside comparison [2]. The technical report carries no individual author list on its title page and is not peer reviewed [2]. The headline "35B beats trillion-parameter models" framing applies to a minority of benchmarks; on most rows the DeepSeek V4 Pro Preview and closed frontier models remain ahead [2]. The self-evolving data pipeline itself, which the team calls broadly applicable with "strong scaling potential", is described but not released [2][4].

## See also

- [Qwen3](https://aiwiki.ai/wiki/qwen3)
- [Absolute Zero](https://aiwiki.ai/wiki/absolute_zero)
- [AlphaEvolve](https://aiwiki.ai/wiki/alphaevolve)
- [Synthetic data](https://aiwiki.ai/wiki/synthetic_data)
- [AI for Science](https://aiwiki.ai/wiki/ai_for_science)
- [Chinese AI](https://aiwiki.ai/wiki/chinese_ai)
- [Open-source AI](https://aiwiki.ai/wiki/open_source_ai)

## References

1. BigBang-V1 model card. Hugging Face, endless-frontier/BigBang-v1, updated August 6, 2026. https://huggingface.co/endless-frontier/BigBang-v1
2. The BigBang Team. "BigBang: Pursuing Open-Ended Intelligence through Self-Evolving Synthesis of Verifiable Frontier Tasks." Technical report (preprint), The Endless Frontier, August 2026. https://endlessfrontier.tech/assets/paper.pdf
3. "BigBang: Pursuing Open-Ended Intelligence through Self-Evolving Synthesis of Verifiable Frontier Tasks." Project site, The Endless Frontier, accessed August 7, 2026. https://endlessfrontier.tech/
4. BigBang-v1 repository. GitHub, endless-frontier/BigBang-v1, created August 3, 2026. https://github.com/endless-frontier/BigBang-v1
5. Hugging Face model API record for endless-frontier/BigBang-v1 (parameters, license, repository dates), accessed August 7, 2026. https://huggingface.co/api/models/endless-frontier/BigBang-v1
6. LICENSE file (Apache License 2.0). Hugging Face, endless-frontier/BigBang-v1. https://huggingface.co/endless-frontier/BigBang-v1/blob/main/LICENSE
7. "35B干赢万亿参数大模型！上交大AI开始给自己造题还能自我迭代了." Synced (机器之心), syndicated at Sina Technology, August 7, 2026. https://finance.sina.com.cn/tech/roll/2026-08-07/doc-inimnvkp4722171.shtml
8. "35B干赢万亿参数大模型！上交大AI开始给自己造题还能自我迭代了." Synced (机器之心), syndicated at NetEase (163.com), August 2026. https://www.163.com/dy/article/L3OFKJNU0511AQHO.html
9. Qwen3.6-35B-A3B model repository. Hugging Face, Qwen, created April 15, 2026. https://huggingface.co/Qwen/Qwen3.6-35B-A3B
10. The Endless Frontier organization page. Hugging Face, accessed August 7, 2026. https://huggingface.co/endless-frontier
11. Adina Yakup (@AdinaYakup). Post on X announcing BigBang-v1, August 7, 2026. https://x.com/AdinaYakup/status/2085674148399898732
12. 上海算法创新研究院 (Institute for Advanced Algorithms Research, Shanghai). Official site, accessed August 7, 2026. https://www.iaar.ac.cn/
13. "Institute for Advanced Algorithms Research, Shanghai." Nature Index institution profile. https://www.nature.com/nature-index/institution-outputs/china/institute-for-advanced-algorithms-research-shanghai/67b8089cf3a59645b4027e79
14. Siheng Chen. Personal academic homepage, Shanghai Jiao Tong University, accessed August 7, 2026. https://siheng-chen.github.io/
15. Yuwen Du, Rui Ye, Shuo Tang, Xinyu Zhu, Yijun Lu, Yuzhu Cai, Siheng Chen. "OpenSeeker: Democratizing Frontier Search Agents by Fully Open-Sourcing Training Data." arXiv:2603.15594, 2026. https://arxiv.org/abs/2603.15594
16. mradermacher/BigBang-v1-GGUF. Hugging Face, community quantization, created August 3, 2026. https://huggingface.co/mradermacher/BigBang-v1-GGUF
17. "Weinan E." Wikipedia, accessed August 7, 2026. https://en.wikipedia.org/wiki/Weinan_E