Ant introduced the multimodal model Ling-3.2-flash with a million-token context window

The Large Language Model Ling team from Ant Group has released Ling-3.1-flash-1019, an updated version of the previously introduced Ling-3.1-flash model, and simultaneously published the underlying pre-trained model, Ling-3.1-flash-base. Both versions are available on the Hugging Face platform under the MIT license.

The architecture of Ling-3.1-flash-1019 retains features of its predecessor: it is a hybrid structure with a linear Mixture-of-Experts mechanism, possessing a total of about 560 billion parameters and approximately 25 billion active parameters per token, along with a context window of one million tokens. According to the model description, the stack includes 63 layers of KDA linear attention and 9 layers of controlled MLA, among which 512 routed experts are present, with 8 activated simultaneously. The main change in this release relates to additional post-training aimed at improving reasoning, tool usage, and multi-step task execution.

According to the third-party artificial intelligence index, Artificial Analysis Intelligence Index v4.3.2, the new version received a score of 44, which is three points higher than the previous release. Other metrics are results claimed by Ant itself. The team reported that its score on the AA-Briefcase test, a comprehensive knowledge test, increased from approximately 1400 to 1524 Elo, and the GDPval-AA metric rose from about 1618 to 1644 Elo. Furthermore, in the area of process automation, AutomationBench-AA improved from 62% to 68%, and scientific problem-solving increased: CritPt rose from 18% to 26%, and SciCode—from 54% to 57%.

The base model represents a more significant addition for developers. Ant asserts that Ling-3.1-flash-base outperformed DeepSeek-V4-Pro-Base and MiMo-V2.5-Pro-Base across 32 out of 47 benchmarks in eight areas, based on proprietary tests covering knowledge, reasoning, code, mathematics, and long context. The weights have been released in BF16 format, allowing teams to conduct further pre-training, domain adaptation, and custom fine-tuning.

Ant positions this release around the concept of personal agents rather than chatbots. The update appeared one day after the Ant LingGuang 2.0 AI assistant application was repositioned as a personal assistant capable of continuing work on tasks after the user concludes the dialogue, utilizing inter-session memory, long-term task planning, and dedicated cloud hosting. Examples provided by the team include tracking weather and bookings before a trip or monitoring product prices against a set budget. The team also stated that it trained an image and language recognition variant, Ling-3.1-flash-VL, which is already used within LingGuang and will be publicly released later.

The recommended serving configurations listed in the model card feature SGLang and vLLM running on eight GPU nodes.

Ant Group has introduced LingGuang 2.0, an updated version of its AI assistant. The new version is equipped with a persistent memory function, allowing it to continue managing tasks even after the user has ended the dialogue.

With appropriate permissions, the assistant can independently determine user needs, track changes, and perform subsequent actions over an extended period, instead of requiring a new command at every stage of work.

Examples of this function's use include checking forgotten recurring subscriptions in Alipay accounts, monitoring prices for flights and hotels after booking, and extracting tasks from school-related notifications in DingTalk groups.

The assistant integrates with services such as Alipay, DingTalk, Feishu, email, and the Xiaomi Mijia smart home platform. Furthermore, it supports external tools via command-line interfaces and MCP. The release of this version followed limited testing conducted earlier this month.

Ant Group released its Ming-3.0 family of speech models on October 20, providing five models for use in speech synthesis, audio understanding, dialogue management, and speech recognition. Developers can apply these models individually or integrate them into voice-based applications.

The text-to-speech models support voice cloning based on sample recordings and also allow for voice design through natural language descriptions. Users can control the delivery of speech by specifying parameters such as emotional tone, emphasis, and tempo.

The audio models are capable of transcribing, summarizing, and analyzing extended recordings while preserving information about speakers, their emotions, and background noise. Ant reported that these models are already integrated into services such as the Afu health application, Alipay search, and the company's internal office tools.