Ling 3.1 Flash
Ling 3.1 Flash is InclusionAI's reasoning model. Each page prints a different end to the free period, and a different context length.
Two weeks in the 30 September report
The Ifeng page carrying IT Home's 30 September 2026 report says 蚂蚁百灵 released Ling-3.1-flash that day. It gives about 560 billion total parameters, with about 25 billion activated for each token, and a context ceiling of 1 million tokens. It says there will be a two-week free trial. During that trial, because of the combined serving cost, the served length is 256K. After the trial ends and the service becomes paid, the report says the team plans to open the 1 million token context, open-source the model at the same time, and keep updating its performance. The same report says the work was aimed at general agents, search, everyday office tasks, and software development, and also at medical, finance, and materials research.
A public search of Hugging Face models under inclusionAI for Ling-3.1 returned none. The org does publish Ling-3.0-flash, a different model, at 124 billion total parameters and 5.1 billion active. On OpenRouter's comparison, that earlier model is listed at US$0.021 per million input tokens and US$0.063 per million output tokens, while the Ling 3.1 Flash row is Free. Both rows on that comparison say 262,144 tokens of context.
Free through 13 October, on two Vercel names
The Vercel changelog, published the same day, says Ling 3.1 Flash from InclusionAI is free on AI Gateway through 13 October 2026. It gives the context on that gateway as 262K tokens. The standard model page and the free model page both name Novita as the provider. Both list a context of 262,144 tokens and a maximum output of 32,768. Both say detailed capability metadata has not been reported for this model. The changelog names two ids.
inclusionai/ling-3.1-flashis free during the promotion and begins billing when the promotion ends. The changelog does not print the price that billing uses.inclusionai/ling-3.1-flash-freestops serving when the promotion ends, instead of starting a charge.
You can call the gateway with an AI Gateway key or with your own provider key. The changelog's setup for Claude Code and for Codex is npx vercel ai-gateway setup, then /model. Codex can also be started with codex --model inclusionai/ling-3.1-flash. fx is listed for macOS and Linux, on x86_64 and arm64. Its install line is curl -fsSL https://fx.sh/setup.sh | bash. With no saved credentials, fx opens a picker to sign in with Vercel or to paste an AI Gateway key, then /model inclusionai/ling-3.1-flash. The same note says the model is meant for coding, multi-step analysis, and agents that use tools.
1 million in the Novita title, 256K in the box
Novita hosts inclusionai/ling-3.1-flash. The title says a 1 million token context window. The spec box on the same page says context length 256K and max output 32K. The price line says input US$0 per million tokens and output US$0 per million tokens. The serverless section also says you pay per token, and that you pay only for usage. The sample request uses https://api.novita.ai/openai and sets max_tokens to 32768. That box marks function calling, reasoning, and the Anthropic API as supported.
OpenRouter says free, and names no end date
OpenRouter lists the model as free, released 2 October 2026. It says one provider hosts it, Novita, and that every request is forwarded there directly. The provider row says input Free and output Free. The FAQ says the price shown is zero, so you are not charged for prompt or completion tokens. That page does not name an end date. It says the context window is 262,144 tokens and a completion can run to 32,768 tokens. It says the model accepts tools and tool_choice. It does not support response_format, so a JSON shape is not enforced.