LFM2.5-8B-A1B — Fastest Local AI Agent on a Laptop? (6 Tests)

追加: 2026-06-02

[00:00:00]What's going on, everybody? Welcome back to the channel. Today, we're talking about a model that on paper should not be possible. Liquid AI just dropped LFM 2.58B A1B, and they're calling it a personal assistant that runs on your laptop. No cloud, no API bills on your machine, and I've already got it running. So, stick around. Here's the headline. It's an 8 billion parameter model that only activates 1 and 1/2 billion at a time.

[00:00:26]It's a mixture of experts. That's what the A1B means. So, you get the knowledge of an 8B model, but the speed closer to a 1B. That is the whole trick, and that's why this thing flies. And it is not your standard transformer. The architecture is a hybrid. 24 layers, 18 of them fast convolution layers, and only six attention layers. That's the Liquid Foundation Model DNA, and it's a big reason they can squeeze this much speed out of the edge. The specs are stacked. A 128,000 token context window, nine languages, native tool calling, explicit chain-of-thought reasoning, trained on 38 trillion tokens, and the benchmark jumps over the last version are wild.

[00:01:05]Instruction following went from 79 to 92. The AIME math score doubled, and the telecom agent benchmark went from 13 to 88. But, I'm going to be straight with you. The most important number for me is hallucinations. Its honesty score went from about seven up to 63. It learned to say, "I don't know." Now, Liquid says it themselves. This is not a coding monster and not for deep knowledge without retrieval. The lane is agents, tool use, and multilingual assistance. So, let's actually test it. Setup was painless. I grabbed the GGUF file from Hugging Face, pointed a model file at it, and ran Ollama create. One thing to note, the raw GGUF ships with an empty chat template, so I dropped in the proper one. A few seconds later, success. The model is sitting on my machine, 5.2 gigs. First test, reasoning. The classic bat and ball trap, where the obvious answer is wrong. Watch it think out loud, set up the equation, and land on 5 cents, not 10. Then it carries the math through for three bets and two balls, correct? Next tool calling the headline feature, I gave it a weather tool and asked about Tokyo. It reasons about which tool to use, then emits a clean Pythonic function call. Feed the result back and it turns it into a natural answer. 11° light rain, bring a jacket.

[00:02:21]That's a full agent loop running locally. Structured output. I handed it a messy sentence about a person and asked for JSON matching a schema, and it nailed every field. Name, role, company, skills, years, remote. This is the stuff that makes it actually useful inside an app. Multilingual, one sentence into Japanese, French, and Arabic. Clean, labeled, and correct. Nine languages out of the box, all from a model small enough to live on your laptop. Now, the honesty test. I ask it to summarize a study that does not exist. A fake paper I made up. A lot of small models will happily invent an answer. This one refused and said it's not aware of that study. That is exactly what you want from a local assistant. And the honest weakness, I asked for a hard algorithm, the median of two sorted arrays in log time. The code looks right, but it's got a subtle bug, so I ran it. Four of five test cases failed. So, yeah, believe Liquid, this is not your coding model, but credit for being up-front about it.

[00:03:21]What about speed? On my RTX 4060 laptop, single stream, it held about 76 tokens a second. Liquid's 18,000 tokens per second figure is an H100 at full concurrency, a totally different beast.

[00:03:34]But 76 on a laptop, fully offline, is genuinely fast. To make it a real daily driver, I wrapped it in a Gradio app, a clean local chat UI talking to Ollama.

[00:03:45]Same model, same speed, but now with a reasoning panel and live tokens per second right in the browser. No cloud anywhere in the loop. So, where does that leave us? The pros, it's private and fully offline. It's fast thanks to that 1.5 B active design. Tool calling and JSON just work. It reasons before it answers. It's multilingual and it runs everywhere. llama.cpp, MLX, vLLM, LM Studio, you name it. And the cons, honestly, it's weak at heavy coding. It leans on retrieval for deep knowledge. The raw GUF needs a proper template before tool calling works in Ollama. The reasoning tokens add a little latency and 8 gigs of VRAM is about the floor. Know the lane and it's brilliant. Bottom line, an 8B model, 1.5 B active, running private on a laptop that holds up as an agentic assistant.

[00:04:38]That's a real shift. If this helped you out, hit subscribe. Link to everything is in the description and I'll see you in the next one.