inclusionAI: Ling 3.0 Flash

Ling 3.0 Flash is a text model from inclusionAI, released by the provider on 2026-07-23. It is a 124B-parameter Mixture-of-Experts (MoE) model with roughly 5.1B parameters activated per token. The design priorities stated by the provider are token efficiency and production-scale agentic inference, which positions it as a model aimed at developers running agent workloads rather than one-off chat.

The model is text in and text out. It accepts text prompts and returns text completions, with a context window of 262144 tokens and a maximum output of 32768 tokens. The large context window means long documents, extended conversation histories, or sizeable tool outputs can be held in a single request, while the output cap bounds how much the model can generate in one response. Reasoning output is supported, so the model can expose its reasoning as part of a response when that is requested.

Those capabilities point to a few practical uses. Long-context work such as summarising or querying large documents fits the 262144-token window. Agentic pipelines fit the stated focus on production-scale agentic inference, especially combined with tool calling and structured output. Because reasoning can be surfaced, the model also suits tasks where a developer wants to inspect intermediate reasoning, such as debugging a multi-step workflow or validating a decision path before acting on it.

You call the model through api14's OpenAI-compatible API using the model id "inclusionai/ling-3.0-flash". The supported request parameters are frequency_penalty, include_reasoning, logit_bias, logprobs, max_tokens, min_p, presence_penalty, reasoning, repetition_penalty, response_format, seed, stop, temperature, tool_choice, tools, top_k, top_logprobs and top_p. Tool calling is available through tools and tool_choice, structured output through response_format, and sampling can be controlled with temperature, top_p, top_k, min_p, the penalty parameters and seed. Log probabilities can be requested with logprobs and top_logprobs, and reasoning behaviour can be tuned with reasoning and include_reasoning.

The main practical constraints are the text-only input and output and the 32768-token output ceiling. If a task depends on image or audio input, this model does not cover it. If a single response needs to exceed the output cap, the work has to be split across requests. Within those bounds, the combination of a 262144-token context, tool calling, structured output and reasoning controls makes it a reasonable default for long-context and agent-style text workloads.