Ant Group launches Ling-3.2-flash, multimodal model with 1 million token context window

Ant Group launched Ling-3.2-flash on November 5, 2026, a natively multimodal model that accepts text, image and video as input. The announcement was made by the Ling project's own team and reported by the Chinese tech site TechNode. The model ships with a 1 million token context window, enough to process at once something like several entire technical books or hours of transcription without needing to break the content into chunks.

Architecture: 560 billion parameters, only 25 billion active

Ling-3.2-flash has 560 billion parameters in total, but activates only 25 billion of them at each inference step. This mixture-of-experts (MoE) design is what makes it possible to run a model of this size without paying the computational cost of using all parameters on every query. In practice, it's the same logic behind other MoE models on the market: the model is large enough to accumulate specialized knowledge, but fast enough not to make the cost per token unviable.

The "flash" name in the model, following a convention that already appears in other AI families, usually indicates the lighter, faster variant of a model line, designed for usage volume rather than maximum latency per response. Ant did not detail in the announcement whether larger variants of the same Ling 3.2 generation exist, nor did it publish comparative benchmarks of speed or quality against other models on the market.

What it's for: vision, agents, 3D and medicine

According to Ant, Ling-3.2-flash was optimized for four specific areas:

• Complex visual understanding (analysis of images with multiple layers of information);
• Multi-step agent tasks (sequences of actions that depend on accumulated context);
• 3D understanding and interaction;
• Multimodal applications in the medical field.

This selection of use cases reveals the audience Ant is targeting: not a generic chat model, but a tool designed for agent pipelines that need to maintain state across many interactions, and for vertical applications (healthcare, computer vision) where processing image and text together is part of the problem, not an accessory feature.

Access compatible with OpenAI and Anthropic, free for two weeks

Ant made Ling-3.2-flash available through two paths: an online chat interface and APIs compatible with OpenAI's and Anthropic's protocols. This means that, in theory, anyone who already has code written against the OpenAI or Anthropic SDK can point their calls to Ant's endpoint by changing little beyond the base URL and the API key, without rewriting prompt logic or response parsing.

Editor's note: the code in this section was not executed in a real environment. Validate before using in production.

# Illustrative example of endpoint swap
# keeping the same OpenAI client already used in the project
from openai import OpenAI

client = OpenAI(
 base_url="https://endpoint-da-ant.exemplo/v1",
 api_key="SUA_CHAVE"
)

response = client.chat.completions.create(
 model="ling-3.2-flash",
 messages=[{"role": "user", "content": "..."}]
)

Access is free during the first two weeks after launch, according to Ant. The company had not disclosed, as of the source's publication, what the billing model will be after that period or the per-token pricing table.

What this changes for agent builders in Brazil

In short: API compatibility lowers the testing barrier for Brazilian teams that have already built products on top of OpenAI's or Anthropic's call format. This matters because switching model providers usually requires rewriting response parsing, function calling format and streaming handling; when the protocol is compatible, this migration cost drops.

For those developing agents that need to maintain memory of long conversations, process extensive documents or analyze sequences of images within the same context, a 1 million token window reduces the need for retrieval architectures (RAG) solely to work around context limits. This doesn't eliminate RAG in every case, but it changes the calculation of when it's worth simplifying the pipeline and feeding more raw context into the model.

The two-week free period is a real no-cost testing window for teams that want to compare response quality, latency and behavior on multimodal tasks before deciding whether it's worth switching providers or simply adding one more option to the product's model router.

What's still unknown

Ant has not published benchmarks comparing Ling-3.2-flash with Western models or with other versions of the Ling family itself, nor has it detailed the per-token price after the free period. The source also contains no information about average latency, requests-per-minute limits, or whether the model (and the data sent to it) is hosted on infrastructure in China, a point Brazilian teams need to check before sending sensitive data to any API, regardless of the model's origin. These details tend to emerge as developers outside China begin testing API access in the coming weeks.

Source

TechNode (China): "Ant launches Ling-3.2-flash multimodal model with million-token context". According to the publication, Ant Group launched Ling-3.2-flash on November 5, a natively multimodal model that accepts text, image and video, with 560 billion total parameters, 25 billion activated per inference step, a 1 million token context window, available via chat interface and APIs compatible with OpenAI's and Anthropic's protocols, with free access during the first two weeks after launch. Read the original report at: https://technode.com/2026/11/06/ant-launches-ling-3-2-flash-multimodal-model-with-million-token-context/.