LiquidAI: LFM2.5-24B-A2B

LiquidAI: LFM2.5-24B-A2B is a text model from LiquidAI. It is a mixture-of-experts reasoning model with 24B total parameters and 2.45B active parameters. The provider describes it as always reasoning before answering, and as built for agents that call tools. It sits in the text category, so it takes text in and returns text out.

The model accepts text input and produces text output, with a context window of 128,000 tokens and a maximum output of 8,192 tokens. Because it is a reasoning model, it works through its reasoning before producing an answer, and the include_reasoning and reasoning parameters let you control how that reasoning is handled in the response. It also supports tool use through the tools and tool_choice parameters, along with structured outputs and response_format for constraining the shape of the output. Sampling behaviour can be tuned with temperature, top_k, min_p, frequency_penalty, presence_penalty, repetition_penalty, seed and stop, and logprobs with top_logprobs are available if you need token-level probabilities.

Those capabilities point to a few practical uses. Tool-calling agents are the clearest fit, since the model is designed to reason and then invoke tools, which suits workflows that need to look things up, call functions or chain steps together. Structured outputs make it a reasonable choice for extracting fields from documents or returning machine-readable results that feed into another system. The 128,000 token context window supports long inputs such as large documents, long conversation histories or sizeable code files, and the reasoning step helps on tasks that benefit from working through a problem before answering, such as multi-step analysis or planning.

You call the model through api14's OpenAI-compatible API using the model id "liquid/lfm-2.5-24b-a2b". Requests follow the OpenAI chat completions format, so existing OpenAI SDKs and clients work with a change of base URL and model id. The supported request parameters are frequency_penalty, include_reasoning, logprobs, max_completion_tokens, max_tokens, min_p, presence_penalty, reasoning, repetition_penalty, response_format, seed, stop, structured_outputs, temperature, tool_choice, tools, top_k and top_logprobs.

Keep the output ceiling in mind: responses are capped at 8,192 tokens, so tasks that need a very long single answer may require splitting the work across calls. The model is text only, so if your input includes images or audio you will need a different model. For straightforward, high-volume text generation where reasoning and tool calls add no value, a simpler model may be cheaper and faster, but for agent-style and structured extraction work this model's feature set is a good match.