Ant released Ling-3.0-flash with 256K context for AI agents
The 124-billion-parameter mixture-of-experts model activates 5.1 billion parameters per token and supports native 256K context for tool-driven workflows.
By Ryan Merket · Published
Why it matters
Ling-3.0-flash gives developers another low-active-compute option for long-context agents and tool calls. Its value will depend on whether Ant's sparse architecture reduces total workflow cost after retries, context processing and failed actions are counted.
On July 24, 2026, Ant Group released Ling-3.0-flash, giving its Ling model family a sparse mixture-of-experts system built for long-running agent workflows, tool calls and high-throughput inference, according to the official release page. Ant's documentation lists 124 billion total parameters, with 5.1 billion activated for each token, a native 256,000-token context window and an option to extend context to 1 million tokens.
The model can switch between reasoning and non-reasoning modes, allowing an application to reserve additional computation for harder prompts. Ant positions Flash for tool calling and production AI-agent workloads.
Sparse computation for agent workloads
Ling-3.0-flash uses a mixture-of-experts architecture, which routes each token through a selected portion of the model rather than activating all 124 billion parameters. Its 5.1 billion active parameters represent the immediate computational workload disclosed for each token, though total serving cost also depends on memory requirements, context length, output volume, hardware and the efficiency of the provider's implementation.
Ant claims Flash can reach inference speeds of up to 1,000 tokens per second and time-to-first-token below 100 milliseconds, according to its model documentation. Those figures are vendor claims and were not independently established in the reviewed research. Performance inside an agent also depends on tool latency, prompt caching, retries and whether the model can complete a workflow without escalating the task to a larger system.
The 256,000-token native context gives developers room to pass large code repositories, document collections or long interaction histories into a single session. Ant says the window can be extended to 1 million tokens, although a larger context can raise memory use and processing cost even when the model activates a relatively small share of its parameters.
Flash is available through the Ant Ling API, OpenRouter and ZenMux, according to Ant's release page and developer materials. The documentation also lists integrations with Claude Code, OpenClaw, Kilo Code and Hermes Agent.
Those channels give developers several ways to test the model without operating inference hardware. Ant has not disclosed Ling-specific customer counts, paid API volume, revenue or model-download figures in the materials reviewed for this story, leaving limited evidence about adoption beyond the published integrations and distribution channels.
Ling operates as an internal Ant research project
Ling is an internal Ant Group research and product initiative. Ant's Ling model documentation describes Ling as an independently developed, open-source general-purpose model series. The research is attributed to a collective Ling Team. A recent Ling and Ring technical report lists Ang Li and Ben Liu first among hundreds of authors, but the public materials do not identify a single project founder or establish that either researcher leads the program.
Ant Group grew out of Alipay, Alibaba's payment service established in 2004, and traces its corporate formation to 2014. The company is headquartered in Hangzhou, China. Ling is funded through Ant's corporate research and product operations; no Ling-specific investors, valuation or capital raised have been disclosed.
Ant's published roadmap places Ling 1.0 in March 2025, Ling 2.0 in October 2025, Ling 2.5 in February 2026 and Ling 2.6 in April 2026, followed by Ling 3.0 in July 2026. Flash therefore extends a model program Ant has revised several times in less than 18 months.
Competition is shifting toward active compute
Ant is competing with model developers that increasingly market sparse architectures and smaller serving footprints. Alibaba's Qwen3-30B-A3B has 30 billion total parameters and activates 3 billion, while also supporting thinking and non-thinking modes. Qwen launched that generation in April 2025 as part of an open-weight lineup spanning several model sizes.
Other companies are approaching lower-cost inference from different directions. Liquid AI offers LFM2 models ranging from 350 million to 2.6 billion parameters for on-device and on-premise use with a non-transformer architecture. IBM's Granite 4.0 includes small and hybrid mixture-of-experts variants aimed at constrained hardware. Fastino, which develops small task-specific enterprise models, raised $17.5 million led by Khosla Ventures in May 2025.
Ling-3.0-flash's commercial case will depend on completed work per dollar. Low active-parameter counts can reduce the cost of individual calls, but agent systems often incur hidden expense through retries, failed tool calls and repeated context processing. Ant has supplied the architecture, access routes and vendor performance targets. Production usage data will determine whether Flash can carry long agent sequences at the speed and cost its specifications imply.