LesenForschungRadarAnlageframework
Anmelden / Registrieren
← Zurück zur Lektüre
NewsThema · 蚂蚁百灵Ling模型Wesentlich1 Min. LesezeitDisconfirmAI Redaktion

Ant Group's Bailing Open-Sources Ling-3.1-flash: 560B Parameters, 1M Context, Targeting High-Cost-Efficiency Agent Tasks

Ant Group's Bailing has open-sourced the Ling-3.1-flash model, featuring ~560B total parameters and a 1M context window, aimed at office automation and complex agent tasks with low cost.

2026-10-15 13:20 GMT+8

Ant Group's Bailing officially open-sourced its latest large model, Ling-3.1-flash, on October 15. The model has approximately 560 billion total parameters, activates about 25 billion parameters per token, and supports a context window of up to 1 million tokens. It is primarily designed for daily office work, long-cycle software development, and complex agent tasks.

According to a report by Zhidongxi, in public tests conducted by third-party evaluation agency Artificial Analysis, Ling-3.1-flash achieved an intelligence index of 41, with an output speed of approximately 210 tokens per second and a mixed calling price of about $0.19 per million tokens. This combination demonstrates a significant cost-performance advantage.

Zhidongxi's editorial team conducted hands-on tests covering financial statement analysis, work information organization, 3D scene construction, and mobile app development. The tests showed that the model could complete the entire workflow from data retrieval and organization to visualization, exhibiting comprehensive execution capabilities in complex long-horizon agent tasks.

It should be noted that these test results are based on specific task scenarios from the media editorial team and are not independent standardized benchmarks. Furthermore, although the model is now open-source, its stability and safety in broader production environments still require further validation by the community.

Quellen

Anmelden / Registrieren
LesenForschungRadarAnlageframework
zhidx.com
↗
Thema

蚂蚁百灵Ling模型

蚂蚁集团旗下百灵团队近期密集更新其开源大模型Ling系列,重点聚焦于面向智能体(Agent)任务的Ling-3.1-flash及其升级版Ling-3.1-flash-1019。该系列模型采用混合线性MoE架构,总参数约560B,支持1M上下…

4 Schritte1 BeiträgeAktualisiert 2026-11-05 17:12

Wie es dazu kam

  1. 2026-09-30蚂蚁百灵发布Ling-3.1-flash模型,总参数约560B,激活参数25B,支持1M上下文。zhidx.com ↗
  2. 2026-10-15智东西报道Ling-3.1-flash正式开源,实测显示其在办公整理和复杂Agent任务中具备较高性价比,输出速度约210 Token/s。zhidx.com ↗
  3. 2026-10-27蚂蚁百灵宣布开源Ling-3.1-flash-1019及基础模型Ling-3.1-flash-base。1019版在Artificial Analysis智能指数中得分为44,较前代提升3分。x.com ↗

Was noch ungeklärt ist

  • 百灵是否会针对Ling-3.1-flash-1019的高Token消耗问题推出更高效的推理优化方案或新的轻量级版本?
  • 随着“灵光”等应用对长程任务需求的增加,蚂蚁集团如何平衡模型能力提升带来的API成本上涨与C端/B端用户的付费意愿?
  • Ling-3.1-flash-base作为基础权重,在社区中的实际微调效果及衍生模型的表现如何?

Auch in diesem Thread

  • Ant Group's Bailing Open-Sources Ling-3.1-flash-1019: Agent Benchmark Scores Rise, Per-Task Cost Up Nearly 50%2026-10-29
Vollständiges Thema lesen →
Recherche fortsetzen →