Agents-A1: 扩展能力边界而非参数规模：用 35B 智能体达到万亿参数级性能

本仓库包含 Hugging Face Transformers 格式的 Agents-A1 模型权重和配置文件。

这些文件兼容 Hugging Face Transformers、vLLM、SGLang 等框架。

---

🔥 新闻

• 2026年7月2日：🔥🔥 基于 Agents-A1，我们发布了一系列量化模型变体，请参阅 Agents-A1 集合。另外，感谢 mlx-community 提供多种规模的量化版本。尝试在你的 Mac 上运行 Agents-A1！

• 2026年6月26日：🔥🔥 我们已开源 Agents-A1 35B-A3B 模型，并同时发布了特定领域的评估代码和技术报告。

---

Agents‑A1 是来自 InternScience 的 35B 混合专家（MoE）智能体模型，旨在扩展覆盖多个领域的异构智能体能力，包括长程搜索、工程、科学研究、指令遵循和工具调用。我们从两个维度探索智能体能力边界的扩展：扩展长程轨迹和扩展异构智能体能力。

在扩展长程轨迹方面，Agents‑A1 借助一套领域驱动的知识-行动基础设施进行训练，该基础设施联合构建行动、观察和验证结果，将智能体的过程转化为可训练目标。在扩展异构智能体能力方面，Agents‑A1 提出了一种三阶段训练范式，用于构建可扩展的通用智能体模型。首先，进行全领域有监督微调，使基础模型对齐广泛的智能体行为。其次，训练领域级教师模型，以捕捉每个领域的专门知识。第三，提出多教师多领域在线策略蒸馏，结合异构感知优化，以提高跨领域的知识迁移效率。

亮点

• 智能体推理：Agents-A1 擅长将复杂任务分解为可执行的子步骤，提前规划，并根据中间结果调整策略。
• 工具使用：原生支持函数调用和工具集成，能够与 API、代码解释器、搜索引擎等外部工具无缝交互。
• 科学与专业推理：处理结合工具的科学推理以及专业知识问答。
• 指令遵循：精确遵循跨多种领域的详细、多约束指令。

我们欢迎开发者和企业集成并试用 Agents-A1，并分享反馈意见。

性能

我们在六个方向上评估了 Agents-A1 在真实世界智能体和研究工作流中的表现——长程搜索、工程任务、科学研究、指令遵循、通用智能体任务和科学智能体任务。尽管处于约 35B 模型级别，Agents-A1 仍能提供与 GPT-5.5、DeepSeek-V4-pro 和 Kimi-K2.6 等前沿大规模系统极具竞争力的性能。它在若干高难度基准上取得了整体最优（SOTA）结果，包括 Seal-0 (56.4)、HiPhO (46.4)、FrontierScience-Olympiad (79.0)、FrontierScience-Research (40.00)、IFBench (80.6) 和 IFEval (94.8)，同时在 BrowseComp (75.5)、XBench-DS-2510 (86.0)、GAIA (96.0)、SciCode (44.3)、带工具的 HLE (47.6) 和 MolBench-bind (56.8) 等多项任务上位居同级别模型之首。这些结果表明，Agents-A1 结合了强大的长程搜索能力、稳健的科学推理能力以及可靠的指令遵循能力，树立了一个高能力、高效率的智能体模型，缩小了与更大规模前沿模型的差距。

🥇 整体最优 🟢 同级别模型（约 35B）最佳

Benchmark	📏 同级别模型 (~35B)	🚀 更大规模模型	⭐ 我们的模型
	Qwen3.5-35B-A3B	Qwen3.6-35B-A3B	Nex-N2-mini
🔍 长程搜索
BrowseComp	61.0	67.93	74.1
XBench-DS-2510	77.0	71.0	82.0
Seal0	41.4	38.74	49.55
GAIA	59.8	78.64	82.52
⚙️ 工程任务
SciCode	37.7	35.8	29.9
MLE-Lite	24.24	34.85	34.85
🧪 科学研究
HLE 带工具	47.4	36.2	32.0
HiPhO	37.0	37.7	38.5
FrontierScience-Olympiad	64.5	60.3	52.0
FrontierScience-Research	2.5	2.9	5.0
📋 指令遵循
IFBench	70.2	64.4	54.08
LongBench-v2	59.0	57.7	59.6
IFEval	91.9	91.3	88.4
🤖 通用智能体任务
τ2-Bench	🟢 81.2	79.0	74.53
VitaBench	31.9	35.6	23.0
🔬 科学智能体任务
MatTools	21.0	15.9	34.1
MolBench-bind	46.0	48.7	51.4

使用方法

SGLang

SGLang 是一个面向大语言模型和视觉语言模型的快速服务框架。

使用 uv 安装 SGLang：

uv venv --python 3.12 --seed --managed-python source .venv/bin/activate

uv pip install sglang

更多细节请参阅其文档。

以下命令在 http://localhost:8000/v1 创建 API 端点：

• 标准版本（1 个 GPU，262K 上下文）：

python -m sglang.launch_server
--model-path InternScience/Agents-A1
--port 8000
--tp-size 1
--mem-fraction-static 0.8
--context-length 262144
--reasoning-parser qwen3

• 工具使用：

python -m sglang.launch_server
--model-path InternScience/Agents-A1
--port 8000
--tp-size 1
--mem-fraction-static 0.8
--context-length 262144
--reasoning-parser qwen3
--tool-call-parser qwen3_coder

vLLM

vLLM 是一个高吞吐量、内存高效的 LLM 推理和服务引擎。

使用 uv 从主分支安装 vLLM：

uv venv --python 3.12 --seed --managed-python source .venv/bin/activate

uv pip install vllm --torch-backend=auto

更多细节请参阅其文档。

以下命令在 http://localhost:8000/v1 创建 API 端点：

• 标准版本（1 个 GPU，262K 上下文）：

vllm serve InternScience/Agents-A1
--port 8000
--tensor-parallel-size 1
--max-model-len 262144
--reasoning-parser qwen3

• 工具调用：

vllm serve InternScience/Agents-A1
--port 8000
--tensor-parallel-size 1
--max-model-len 262144
--reasoning-parser qwen3
--enable-auto-tool-choice
--tool-call-parser qwen3_coder

• 纯文本（跳过视觉编码器以释放 KV 缓存内存）：

vllm serve InternScience/Agents-A1
--port 8000
--tensor-parallel-size 1
--max-model-len 262144
--reasoning-parser qwen3
--language-model-only

推荐采样参数

为获得最佳生成质量，推荐以下采样参数：

• temperature: 0.85
• top_p: 0.95
• top_k: 20
• min_p: 0.0
• presence_penalty: 1.1
• repetition_penalty: 1.0

智能体能力评估

为了向社区提供统一的智能体评估代码库以便公平比较，我们还开源了一个评估框架，用于评估智能体模型的核心能力，包括工具使用和多步推理。评估代码包含在本仓库的 Agents-A1/evaluation 中。

我们使用该框架在标准化和可复现的设置下评估了已发布的模型。具体而言，模型在一组面向智能体的任务上进行测试，这些任务要求模型理解用户目标、分解复杂指令、必要时与工具或环境交互，并产生最终结果。模型卡 中报告的评价结果，正是使用上述开源框架生成的，以便用户可以复现实验，在相同协议下比较其他模型，并进一步为新的智能体场景扩展基准。（请注意： 为确保公平比较，我们报告原始技术报告中的基准结果。如果某模型未报告相应基准结果，我们则采用与我们的模型相同的评估协议对其进行评估。）

有关详细的评估脚本、任务定义、指标及复现说明，请参阅评估代码库。

引用

如果您觉得我们的工作有帮助，欢迎引用。

@misc{bai2026scalinghorizonparametersreaching, title={Scaling the Horizon, Not the Parameters: Reaching Trillion-Parameter Performance with a 35B Agent}, author={Lei Bai and Zongsheng Cao and Yang Chen and Zhiyao Cui and Shangheng Du and Yue Fan and Shiyang Feng and Zijie Guo and Haonan He and Liang He and Xiaohan He and Shuyue Hu and Yusong Hu and Songtao Huang and Yichen Jiang and Hao Li and Xin Li and Dahua Lin and Weihao Lin and Fenghua Ling and Dongrui Liu and Zhuo Liu and Runmin Ma and Chunjiang Mu and Haoyang Peng and Tianshuo Peng and Jinxin Shi and Luohe Shi and Boyuan Sun and Zelin Tan and Shengji Tang and Qianyi Wang and Yiming Wu and Yi Xie and Xiangchao Yan and Jingqi Ye and Peng Ye and Fangchen Yu and Jiakang Yuan and Bihao Zhan and Bo Zhang and Chen Zhang and Shufei Zhang and Shuaiyu Zhang and Wenlong Zhang and Yiqun Zhang and Junpeng Zhao and Zhijie Zhong and Bowen Zhou and Yuhao Zhou}, year={2026}, eprint={2606.30616}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2606.30616}, }

Agents-A1: Scaling the Horizon, Not the Parameters: Reaching Trillion-Parameter Performance with a 35B Agent

This repository contains model weights and configuration files for Agents-A1 in the Hugging Face Transformers format.

These artifacts are compatible with Hugging Face Transformers, vLLM, SGLang, etc.

---

🔥 News

• 2026.7.2 : 🔥🔥 Based on Agents-A1, we have released a series of quantized model variants. Please refer to the Agents-A1 collection. Besides, we’d like to thank the mlx-community for providing quantized versions at multiple scales. Try running Agents-A1 on your Mac!

• 2026.6.26 : 🔥🔥 We have open-sourced the Agents-A1 35B-A3B model, along with the evaluation code for selected domains and the technical report.

---

Agents‑A1 is a 35B Mixture‑of‑Experts agentic model from InternScience, built to scale heterogeneous agentic abilities across multiple domains including Long‑horizon Search, Engineering, Scientific Research, Instruction Following, and Tool-calling. We investigate agent-horizon scaling from two perspectives: scaling long-horizon trajectories and scaling heterogeneous agent abilities.

From the scaling of long-horizon trajectories, Agents‑A1 is trained with the assistance of a domain-grounded knowledge-action infrastructure that jointly constructs actions, observations, and verifier outcomes, turning the agent's process into a trainable target. From the scaling of heterogeneous agent abilities, Agents‑A1 presents a three-stage training paradigm for building scalable general-purpose agentic model. First, we perform full-domain supervised fine-tuning to align the base model with broad agentic behaviors. Second, we train domain-level teacher models to capture specialized expertise in each domain. Third, we propose multi-teacher multi-domain on-policy distillation with heterogeneity-aware optimization to improve knowledge transfer efficiency across different domains.

Highlights

• Agentic Reasoning : Agents-A1 excels at decomposing complex tasks into executable sub-steps, planning ahead, and adapting its strategy based on intermediate results.
• Tool Use : Natively supports function calling and tool integration, enabling seamless interaction with APIs, code interpreters, search engines, and other external tools.
• Scientific and Professional Reasoning : Handles tool-integrated scientific reasoning and professional knowledge question answering.
• Instruction Following : Precisely follows detailed, multi-constraint instructions across diverse domains.

We welcome developers and enterprises to integrate and try Agents-A1 and share their feedback.

Performance

We evaluate Agents-A1 in real-world agentic and research-oriented workflows across six directions — long-horizon search, engineering tasks, scientific research, instruction following, general agentic tasks, and scientific agentic tasks. Despite operating in the ~35B model class, Agents-A1 delivers highly competitive performance against frontier-scale systems such as GPT-5.5, DeepSeek-V4-pro, and Kimi-K2.6. It achieves overall SOTA results on several challenging benchmarks, including Seal-0 (56.4), HiPhO (46.4), FrontierScience-Olympiad (79.0), FrontierScience-Research (40.00), IFBench (80.6), and IFEval (94.8), while also ranking as the best among comparable models on a broad range of tasks such as BrowseComp (75.5), XBench-DS-2510 (86.0), GAIA (96.0), SciCode (44.3), HLE with tools (47.6), and MolBench-bind (56.8). These results show that Agents-A1 combines strong long-horizon search ability, robust scientific reasoning, and reliable instruction following, establishing it as a highly capable and efficient agentic model that narrows the gap with much larger frontier models.

🥇 Overall SOTA 🟢 Best Among Comparable Models (~35B)

Benchmark	📏 Comparable Models (~35B)	🚀 Larger-scale Models	⭐ Ours
Qwen3.5-35B-A3B	Qwen3.6-35B-A3B	Nex-N2-mini	Step-3.5-Flash
🔍 Long-horizon Search
BrowseComp	61.0	67.93	74.1
XBench-DS-2510	77.0	71.0	82.0
Seal0	41.4	38.74	49.55
GAIA	59.8	78.64	82.52
⚙️ Engineering Tasks
SciCode	37.7	35.8	29.9
MLE-Lite	24.24	34.85	34.85
🧪 Scientific Research
HLE w/ tools	47.4	36.2	32.0
HiPhO	37.0	37.7	38.5
FrontierScience-Olympiad	64.5	60.3	52.0
FrontierScience-Research	2.5	2.9	5.0
📋 Instruction Following
IFBench	70.2	64.4	54.08
LongBench-v2	59.0	57.7	59.6
IFEval	91.9	91.3	88.4
🤖 General Agentic Tasks
τ2-Bench	🟢 81.2	79.0	74.53
VitaBench	31.9	35.6	23.0
🔬 Scientific Agentic Tasks
MatTools	21.0	15.9	34.1
MolBench-bind	46.0	48.7	51.4

Usage

SGLang

SGLang is a fast serving framework for large language models and vision language models.

Install SGLang with uv:

uv venv --python 3.12 --seed --managed-python
source .venv/bin/activate

uv pip install sglang

See its documentation for more details.

The following commands create API endpoints at http://localhost:8000/v1:

• Standard Version (1 GPUs, 262K context):

  python -m sglang.launch_server \
    --model-path InternScience/Agents-A1 \
    --port 8000 \
    --tp-size 1 \
    --mem-fraction-static 0.8 \
    --context-length 262144 \
    --reasoning-parser qwen3
  

• Tool Use :

  python -m sglang.launch_server \
    --model-path InternScience/Agents-A1 \
    --port 8000 \
    --tp-size 1 \
    --mem-fraction-static 0.8 \
    --context-length 262144 \
    --reasoning-parser qwen3 \
    --tool-call-parser qwen3_coder
  

vLLM

vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs.

Install vLLM from the main branch via uv:

uv venv --python 3.12 --seed --managed-python
source .venv/bin/activate

uv pip install vllm --torch-backend=auto

See its documentation for more details.

The following commands create API endpoints at http://localhost:8000/v1:

• Standard Version (1 GPUs, 262K context):

  vllm serve InternScience/Agents-A1 \
    --port 8000 \
    --tensor-parallel-size 1 \
    --max-model-len 262144 \
    --reasoning-parser qwen3
  

• Tool Call :

  vllm serve InternScience/Agents-A1 \
    --port 8000 \
    --tensor-parallel-size 1 \
    --max-model-len 262144 \
    --reasoning-parser qwen3 \
    --enable-auto-tool-choice \
    --tool-call-parser qwen3_coder
  

• Text-Only (skips vision encoder to free KV cache memory):

  vllm serve InternScience/Agents-A1 \
    --port 8000 \
    --tensor-parallel-size 1 \
    --max-model-len 262144 \
    --reasoning-parser qwen3 \
    --language-model-only
  

Recommended Sampling Parameters

For the best generation quality, we recommend the following sampling parameters:

• temperature: 0.85
• top_p: 0.95
• top_k: 20
• min_p: 0.0
• presence_penalty: 1.1
• repetition_penalty: 1.0

Agent Capability Evaluation

To provide the community with a unified agent evaluation codebase for fair comparison, we have also open-sourced an evaluation framework for assessing agentic models across core capabilities, including tool use and multi-step reasoning. The evaluation code is included in the Agents-A1/evaluation of this repository.

We use this framework to evaluate the released model under a standardized and reproducible setting. Specifically, the model is tested on a set of agent-oriented tasks that require it to understand user goals, decompose complex instructions, interact with tools or environments when necessary, and produce final results. The evaluation results reported in Model Card are generated using the open-source framework above, so that users can reproduce the experiments, compare other models under the same protocol, and further extend the benchmark for new agent scenarios. (Note that: To ensure a fair comparison, we report the benchmark results from their original technical reports. If a model does not report the corresponding benchmark results, we evaluate it using the same evaluation protocol as our model.)

For detailed evaluation scripts, task definitions, metrics, and reproduction instructions, please refer to the evaluation codebase.

Citation

If you find our work helpful, feel free to give us a cite.

@misc{bai2026scalinghorizonparametersreaching,
      title={Scaling the Horizon, Not the Parameters: Reaching Trillion-Parameter Performance with a 35B Agent}, 
      author={Lei Bai and Zongsheng Cao and Yang Chen and Zhiyao Cui and Shangheng Du and Yue Fan and Shiyang Feng and Zijie Guo and Haonan He and Liang He and Xiaohan He and Shuyue Hu and Yusong Hu and Songtao Huang and Yichen Jiang and Hao Li and Xin Li and Dahua Lin and Weihao Lin and Fenghua Ling and Dongrui Liu and Zhuo Liu and Runmin Ma and Chunjiang Mu and Haoyang Peng and Tianshuo Peng and Jinxin Shi and Luohe Shi and Boyuan Sun and Zelin Tan and Shengji Tang and Qianyi Wang and Yiming Wu and Yi Xie and Xiangchao Yan and Jingqi Ye and Peng Ye and Fangchen Yu and Jiakang Yuan and Bihao Zhan and Bo Zhang and Chen Zhang and Shufei Zhang and Shuaiyu Zhang and Wenlong Zhang and Yiqun Zhang and Junpeng Zhao and Zhijie Zhong and Bowen Zhou and Yuhao Zhou},
      year={2026},
      eprint={2606.30616},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2606.30616}, 
}