LFM2.5 24B joined the picker the day Liquid AI released it

Liquid AI announced LFM2.5-24B-A2B on October 13, 2026. We added it to the tryingopen picker the same day, where it appears as LFM2.5 24B. As of October 14, 2026, it is the newest model on the site.

Its first day on the site also turned up a bug: when it tried to search the web, its replies came back empty. We fixed it on October 14.

What Liquid published

Liquid's announcement, LFM2.5-24B-A2B: Built for Fast On-Device Agents, calls it "a 24B-parameter reasoning model with just 2B active parameters". The Hugging Face model card gives the finer figures.

A mixture of experts holds many small expert blocks and sends each token through only a few of them. Here that means 4 of the 64 per block. The card puts the active share at 2.45B of 24.0B, about a tenth of the weights for any one token. That small active share is where the speed comes from. What total and active parameters tell you about a model explains what the two numbers mean for memory and speed.

The card calls it a "pure reasoning model that always thinks before it answers". Expect a thinking pass before every reply, even a short one.

By Liquid's own measurements, the model is competitive with Qwen3.6-35B-A3B on five of eight agentic and tool-use benchmarks and runs at 184 tokens per second on an Apple M5 Max.

The weights are on Hugging Face, and Liquid lists day-one support for llama.cpp, MLX, vLLM, SGLang and ONNX. The LFM Open License v1.0 allows commercial use for companies under $10 million in annual revenue. Larger companies need a separate license from Liquid, and research use has no revenue limit. What counts as an open model, and what the licenses allow compares it with the other licenses in the picker.

How we serve it

LFM2.5 24B runs in the cloud, reached through OpenRouter with a zero-data-retention requirement on every message. As of October 2026, Liquid's own endpoint is the only one OpenRouter has for the model, and that endpoint accepts the requirement. If the endpoint ever stopped accepting it, your message would fail with an error.

A few things follow from that single endpoint and from the model itself:

• Replies stop at 8,192 tokens. Liquid's endpoint sets that cap.
• Text only. For pictures, pick a model that reads them, such as LFM2.5-VL 3B. You can still attach a PDF on the ordinary page, since OpenRouter's file parser turns it into text first (that step sits outside zero retention, and Self Chat refuses PDFs).
• Tools work. It can make files with the page's file tool, and those files open in a preview beside the chat. Sending pictures and PDFs, and getting files back covers previews.
• It is in Self Chat. LFM2.5 24B is one of the models Self Chat offers, since it meets that mode's rule of a zero-retention endpoint reached directly through OpenRouter.

In the picker it sits under "Quick and light", and its row reads "Quick thinking and agent work". Liquid calls it "the largest and final model in the LFM2.5 family". It keeps the architecture of the earlier LFM2-24B-A2B and adds a longer context, a bigger vocabulary and reasoning-only training. Its details panel lists a 128k context. That is the 128,000 tokens Liquid's endpoint on OpenRouter serves, a little under the model's full 131,072.

Liquid recommends pairing it with search

The model card is direct about one limit. It says the model is "not recommended for knowledge-heavy tasks on its own" and suggests pairing it with retrieval or web search. On tryingopen, web search is that pairing.

On the ordinary page, models that can search get OpenRouter's web-search server tool, which OpenRouter runs itself when the model asks. The system prompt normally carries a line telling the model it can search. How web search works on tryingopen has the full picture.

What went wrong on day one

With that line in its system prompt, LFM2.5 24B called a function named "webSearch" as if it were one of the page's own tools. The page has no function by that name, so the call went unanswered and the turn ended with an empty reply.

That happened in 4 of 4 tries, including when LFM2.5 24B served as the cloud fallback for our default pick, the small model on our Jetson boards. There, a question the d1 router had sent to the cloud came back empty.

On October 14 we stopped naming the tool in this model's system prompt and kept the tool attached to its requests. Left to find the tool on its own, LFM2.5 24B now searches through OpenRouter's server tool and cites its sources like the other models. Every other model that gets the search tool still gets that line, and none of them has had trouble with it.

One privacy note applies to search on every model: the search query goes to a search provider through OpenRouter, and that step sits outside the zero-retention requirement. In Self Chat, search stays off until you switch it on in the composer. What zero data retention covers, and where it stops goes through the details.

The three Liquid models on tryingopen

LFM2.5 24B is the third Liquid AI model on tryingopen, and the only one that runs in the cloud.

You can put all three in one chat. Keep the default LFM2.5-VL 3B pick and choose LFM2.5 24B in the "Cloud fallback" picker. d1 then reads each message, the small model keeps the quick ones, and LFM2.5 24B takes the rest. The page address carries that setup, so you can save it as a preset or send it as a link. Sharing a chat, a model setup or a preset as a link shows how.

For the two Liquid models on our boards, see Our default model runs on a palm-sized Jetson Orin Nano and How the d1 router picks between our small model and the cloud. For everything else added recently, see What's new in the picker since September.