Liquid AI LFM2.5-24B-A2B: a 24B open-weight MoE with 2B active parameters, its license and what it needs to run

On October 13, 2026, Liquid AI released LFM2.5-24B-A2B, a 24-billion-parameter reasoning model whose weights you can download from Hugging Face. It's a mixture-of-experts (MoE) model: only about 2.45 billion parameters do work for each token, which is why Liquid pitches it for "fast, private agents on-device." It joins a crowded field of open models in the 20B–35B range, and it comes with a license that matters more than usual: it's free for small companies, not for big ones.

This story is based on Liquid AI's launch post, its Hugging Face model cards, its documentation and the LFM license text, checked October 14, 2026.

The short version

• Size: 24B total parameters, about 2.45B active per token (Liquid rounds to "2B"); 64 experts per MoE layer, 4 used per token.
• Context: 131,072 tokens; Liquid says it extended the window from 32,768.
• Type: reasoning-only. It always writes a chain of thought before answering.
• License: LFM Open License v1.0. Commercial use isn't licensed for companies with $10 million or more in annual revenue.
• API: $0.10 input, $0.05 cached input, $1.00 output per million tokens on Liquid's console.
• Speed (vendor-reported): 184 tokens/s on an Apple M5 Max; 114 on an AMD Ryzen AI Max+ 395.

What is LFM2.5-24B-A2B?

Liquid AI calls it the largest and final model in its LFM2.5 family. Unlike most open models, it isn't a standard transformer. According to the model card, its 40 layers are 30 "double-gated short convolution" blocks and 10 grouped-query attention blocks. Fewer attention layers mean a smaller attention cache as conversations grow, one reason Liquid's models are built for constrained hardware.

Last verified: October 14, 2026 (Hugging Face model card and Liquid docs)

Liquid says it was post-trained for tool calling and multi-step agent work, partly through reinforcement learning inside agent harnesses such as OpenCode and Hermes Agent. Its blog makes an unusual admission: during training, models tried to "reward hack" by searching online or in git history for answers. Liquid says it tightened its sandboxes and that all reported results come from sealed environments.

Is LFM2.5-24B-A2B open source?

It's open-weight, not open source in the strict sense. The weights use the LFM Open License v1.0, which Liquid says is based on Apache 2.0 with one big change. The license defines a "Threshold" as "annual revenue of 10 million United States dollars ($10,000,000) or more," and says commercial use by a legal entity over that threshold "is not licensed." Liquid's FAQ points larger companies to a separate commercial license (sales@liquid.ai).

What that means in practice:

• Startups and individuals below $10 million in annual revenue get Apache-style rights: use, modify, fine-tune and redistribute, with notices kept.
• Larger companies can't use it commercially without a deal with Liquid.
• Non-profits (501(c)(3) or equivalent) are exempt from the threshold for non-commercial or research use.
• Fine-tunes can carry your own terms for your changes, but the derivative is still governed by the LFM license, so the revenue gate follows it.

That puts Liquid in the revenue-gated middle tier we described in our open-weight licenses comparison, alongside Mistral's and Qwen's caps, with a lower bar: Liquid counts annual revenue of $10 million, while Mistral Medium 3.5's cap is $20 million in monthly revenue. Read the license before you build a product on it, and again if you might be acquired.

What hardware do you need?

All the weights have to sit in memory, even though only about a tenth are active per token. That's the general MoE rule in our guide to small models for laptops: it computes like a small model but needs the memory of a big one.

Last verified: October 14, 2026 (LiquidAI/LFM2.5-24B-A2B-GGUF on Hugging Face)

Add room for the attention cache and your operating system on top of the file size. As a rough guide, a 4-bit build is a fit for a 32 GB laptop, not a 16 GB one.

It runs on llama.cpp, MLX (Apple Silicon; the docs say mlx-lm 0.32.0 or later, because earlier versions route experts incorrectly), vLLM, SGLang, Transformers, LM Studio and ONNX Runtime. The GGUF page also gives an Ollama command: ollama run hf.co/LiquidAI/LFM2.5-24B-A2B-GGUF:Q4_K_M. If you haven't used Ollama, start with our Ollama walkthrough. Because it's a reasoning model, budget extra output tokens for its thinking.

How fast is it, and how good? (vendor-reported)

All numbers here are vendor-reported, published by Liquid AI on October 13, 2026, and not independently reproduced.

• On-device speed: 184 tokens per second decoding on an Apple M5 Max and 114 on an AMD Ryzen AI Max+ 395, using llama.cpp with the Q4_K_M build. Liquid says that's up to 1.7 times faster than comparable MoE models. The model card says these runs used a 4K-token input context on Liquid's own test infrastructure; batch size isn't given.
• Server throughput: more than 21,000 output tokens per second on one Nvidia B300 at high concurrency (SGLang, BF16), a total across many users, not per-user speed.

Liquid ran every model here with vLLM, with settings varying by test. By its own numbers, the model is strong at tool use and instruction following and trails rivals with more active parameters on science and coding; Liquid suggests pairing it with retrieval for knowledge-heavy work.

What it means for you

For individuals and small teams, this is a new long-context local agent option, and the $0.10/$1.00 API is a cheap way to try it. At a company with $10 million or more in annual revenue, check the license first. For when self-hosting actually saves money, see frontier APIs vs open models.

What's confirmed, reported and uncertain

• Confirmed (Liquid AI): the October 13 release; architecture and parameter counts; 131K context; languages; API prices; file sizes; supported runtimes; the license text and its $10 million annual-revenue threshold.
• Vendor-reported: all speed and benchmark figures; "up to 1.7x faster"; lowest memory use among compared MoEs.
• Uncertain: independent quality and speed results; how Liquid prices commercial licenses above the threshold; the card's 33T vs 36T training-token figures.