Liquid AI released LFM2.5-24B-A2B on 13 October 2026, a model built to run AI agents quickly and privately on a laptop or other personal device rather than in a distant data centre. The company announced it on its blog and in a post on Hugging Face, and says it is the largest and final model in its LFM2.5 family [1][2].
Big model, small working part
Liquid AI describes the model as having 24 billion parameters, with just 2 billion “active” at a time [1]. Parameters are the numbers a model learns during training; more of them generally means more knowledge and skill.
The model uses a design called mixture-of-experts. Instead of one giant block that runs in full for every word, the model is split into many smaller specialist parts, or “experts”. For each piece of text it handles, only a few of those experts switch on. So the total parameter count (24 billion) shows how much the model knows, while the active count (2 billion) shows how much work it does for each word. That is why a model this large can still run quickly on a single machine.
Liquid AI says the LFM2.5 family now runs from LFM2.5-230M, a tiny model, up to this one [1].
How it compares, according to Liquid AI
Liquid AI says the model is competitive with Alibaba’s Qwen3.6-35B-A3B, which has about 50% more total and active parameters, on five of eight agentic and tool-use benchmarks [1]. These tests check how well a model handles agent tasks, such as calling software tools and completing multi-step jobs.
The company’s two headline figures:
- LiveClaw: 48.09 for LFM2.5-24B-A2B, against 38.19 for Qwen3.6-35B-A3B [1].
- AutomationBench: 27.69 against 8.72 [1].
Liquid AI’s own results table on Hugging Face also shows the Qwen model scoring higher on several tests, including PinchBench, WebArena Verified, SWE-Bench Verified for coding and the GPQA Diamond science test [2]. All of these scores are Liquid AI’s own and have not been independently checked.
Speed
Because only 2 billion parameters work on each word, Liquid AI says the model runs up to 1.7 times faster than comparable mixture-of-experts models on devices [1]. The company reports 184 tokens per second on an Apple M5 Max laptop chip and up to 21,000 output tokens per second on a single NVIDIA B300 data-centre GPU [1]. A token is a small chunk of text, often part of a word. On Hugging Face, Liquid AI adds a figure of 114 tokens per second on AMD’s Ryzen AI Max+ 395 [2]. These speeds are also company claims.
What changed from the previous version
The model builds on Liquid AI’s earlier LFM2-24B-A2B. According to the company [1]:
- Longer memory: the context window, meaning how much text the model can consider at once, grew from 32,768 to 131,072 tokens.
- Bigger vocabulary: the set of text chunks it recognises doubled from 65,000 to 128,000, which Liquid AI says helps with languages that do not use the Latin alphabet.
- More training: pre-training was scaled up from 17 trillion to 33 trillion tokens of text.
It always thinks first
Unlike its predecessor, the new model is “reasoning-only”, Liquid AI says: it always writes out a step-by-step chain of thought before giving its final answer [1]. The company argues this makes sense because the model is fast, so the extra thinking costs little time.
Open weights and pricing
Both the base model and the finished, post-trained version are open-weight, Liquid AI says, and available now on Hugging Face and in the company’s online Playground [1]. Open-weight means the trained model itself, its learned numbers, can be downloaded for free. People can run it on their own hardware, adapt it and build on it, without sending their data to Liquid AI. Liquid AI says it works with common tools including llama.cpp, MLX, vLLM, SGLang and ONNX [2].
For people who would rather use it over the internet, Liquid AI’s paid service, Liquid Console, charges $0.10 per million input tokens ($0.05 for cached input) and $1.00 per million output tokens, and gives new users $2 of credit [1].
What this means
- Agents can stay on your machine. A fast, open model aimed at tool use lets developers build assistants that keep data local.
- The comparisons are the company’s. Liquid AI picked the tests and ran them. The rival Qwen model still leads on several, including coding.
- Small active size is the selling point. Mixture-of-experts lets a 24-billion-parameter model behave, for speed, more like a 2-billion one.
Sources
- Liquid AI, “LFM2.5-24B-A2B: Built for Fast On-Device Agents”, 13 October 2026. https://www.liquid.ai/blog/lfm2-5-24b-a2b
- Liquid AI on Hugging Face, “Run fast, private agents on-device with LFM2.5-24B-A2B”, 13 October 2026. https://huggingface.co/blog/LiquidAI/lfm25-24b-a2b
Related stories
- Small Open Models That Answer in One Pass: Liquid AI’s d1 and Perplexity’s pplx-embed-v2-late
- Same Model, Different Score: Training AI Inside the Coding Tools People Actually Use
- NVIDIA Readies Its RTX Spark Laptops for Local AI Agents With a New Open Model, Nemotron 3.5 Bolt
- Ollama Just Made Apple Silicon the Fastest Platform for Local AI
