Ling-3.0-flash — Model availability¶

Development line: project:ling-3-0-flash · thread model-availability
Last event: 2026-07-27 · 1 dated since 2026-07-27 · Researched: 2026-09-08 · confidence: high

What it is¶

Ling-3.0-flash is an Ant Group text-generation MoE for teams building coding, tool-using, and long-context agents.

• Native hybrid-linear MoE architecture with reasoning and tool calling.
• OpenRouter route exposes a 262,144-token context and up to 32,768 output tokens.

124B total parameters, about 5.1B active per token; the OpenRouter route is text-only. It works for agent workloads, but we must keep JSON validation and tool execution in the application.

Development line¶

• 2026-07-27 — Ling-3.0-flash received a dated OpenRouter model reference. On 2026-07-27, an entry linked to OpenRouter's public model page for InclusionAI Ling-3.0-flash. That provides a dated public reference for the model, but does not establish its launch date, capabilities, or current service availability.

What changed¶

• 2026-07-23 — Ant Ling v2.0.0 made Ling-3.0-flash available in chat and through OpenAI- and Anthropic-compatible APIs, with a deep-thinking toggle.
• 2026-07-24 — The formal release announcement added a limited free-access period on Ant Ling and OpenRouter.
• 2026-07-27 — Ant Ling v2.0.1 added the Explore module; it was not a Ling-3.0-flash model release or provider-listing change.
• 2026-08-02 — The official base-weight repository was initialized and populated.
• 2026-08-04 — Official FP8 and INT4 weight repositories were initialized and populated.

How to use this¶

1. Send an OpenAI-compatible Chat Completions request to OpenRouter with the exact model ID inclusionai/ling-3.0-flash, a bearer key, and messages. — https://openrouter.ai/docs/quickstart
2. Keep prompt plus completion within the current 262,144-token context; cap a completion at 32,768 tokens or lower for the workload. — https://openrouter.ai/inclusionai/ling-3.0-flash
3. For tool use, send tools and tool_choice, execute the returned call in the client, append its result, and make the next model request. — https://openrouter.ai/docs/guides/features/tool-calling
4. During a production pilot, send X-OpenRouter-Metadata: enabled and retain the returned routing metadata to see the selected provider and fallback attempts. — https://openrouter.ai/docs/guides/features/router-metadata

Best practices¶

• Treat OpenRouter price, availability, and provider figures as current observations rather than a fixed model property; recheck them before making a budget or latency commitment. — https://openrouter.ai/inclusionai/ling-3.0-flash
• Do not rely on response_format for strict JSON on the current OpenRouter route: its model page says JSON is not enforced. Validate parsed output in the application. — https://openrouter.ai/inclusionai/ling-3.0-flash
• Keep tool execution client-controlled: the model proposes a call, while the application performs it and returns its result. — https://openrouter.ai/docs/guides/features/tool-calling
• For a strict JSON-schema workload, use only an endpoint that currently advertises structured-output support and require supported parameters in routing preferences. — https://openrouter.ai/docs/guides/features/structured-outputs
• For self-hosting, start from the publisher's sampling recipe (temperature=0.6, top_p=0.95, top_k=20) and its runtime-specific parser configuration rather than copying it blindly to another gateway. — https://huggingface.co/inclusionAI/Ling-3.0-flash

Superseded by this¶

• 2026-07-27 — Treating the OpenRouter-linked event date as the Ling-3.0-flash launch date; the official changelog puts the model launch on 2026-07-23 and the July 27 change is Explore.
• 2026-08-03 — Launch-period free API access; the formal announcement set its end at 23:00 Beijing time, so it is not current pricing guidance.

Still unknown¶

• No archived first-listing evidence establishes that OpenRouter first exposed this model on 2026-07-27; the dated link proves only an observation on that date.
• The current OpenRouter provider set, prices, routing, and feature matrix are volatile and were not projected backward into the July event.
• Ling-3.0-flash-VL is a separately named multimodal model; the reviewed sources do not establish that it is a checkpoint update of this text-only model.

Sources¶

Agent brief¶

• Subject: project:ling-3-0-flash, thread model-availability, 1 dated events 2026-07-27 → 2026-07-27.
• Practical note: See the sourced usage and practice sections above, including their limits.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.