A 24B agent model that only wakes up 2B parameters at a time 🤖

Running an AI agent all day costs a lot of tokens. Before it does anything useful, a good agent plans, calls a tool, reads the result and thinks again. If every one of those reasoning tokens goes through a big model in the cloud, you pay for it in money, in latency and in privacy. That's why one release from last week caught my eye: Liquid AI's LFM2.5-24B-A2B, an open-weight reasoning model built specifically for agents that run on your own hardware.

What it is

LFM2.5-24B-A2B has 24 billion parameters in total, but it's a mixture-of-experts (MoE) model: for each token only about 2 billion of them are active. Each MoE block has 64 experts and routes every token to 4 of them. The architecture mixes gated short convolutions with grouped-query attention, the same hybrid as its predecessor LFM2-24B-A2B. New in this version is a context window of 131K tokens (up from 32K), and the vocabulary has doubled from 65K to 128K tokens, which helps with non-Latin scripts.

It's a reasoning-only model: it always writes out a chain of thought before it answers. Normally that makes a model slow and expensive. Here it doesn't hurt much, because each of those thinking tokens only touches 2B active parameters.

How they trained it

The training recipe is the part I find most interesting. Pre-training doubled from roughly 17T to roughly 33T tokens, followed by mid-training on reasoning, code, math and long-context data. Tool calls and multi-turn trajectories show up throughout. Post-training then has five stages: supervised fine-tuning, specializing teacher models, two rounds of on-policy distillation from those teachers (first on non-agentic tasks, then on agentic ones), and finally reinforcement learning inside real agent harnesses such as OpenClaw, Hermes Agent, Pi and OpenCode. So the model isn't only trained to call tools in theory. It practised in the same environments people actually use.

The numbers, honestly

Liquid AI compared it on 16 benchmarks with Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning, all of which are bigger. The strengths are clear:

  • LiveClaw: 48.09, compared with 38.19 for Qwen3.6-35B-A3B
  • AutomationBench: 27.69, more than three times the next-best model (8.72)
  • IFEval: 96.75, the best of the four, and it also leads on IFBench and Multi-IF

It's also fast: 184 tokens per second on an Apple M5 Max, 114 tok/s on a Ryzen AI Max+ 395, and over 21K output tokens per second at high concurrency on a single NVIDIA B300.

What I like is that the team is upfront about the weak spots. The bigger models still win on knowledge-heavy tests like GPQA Diamond (74.42 compared with 83.69 for Qwen) and on coding: SWE-Bench Verified is 51.0 compared with 66.2. Their own advice is to pair it with retrieval or web search for knowledge-heavy work. That's a sensible way to think about small models: let them reason and use tools, and let the facts come from outside.

Why this matters to me

For my homelab this is exactly the right direction. An agent that can run around the clock on a single machine, call tools reliably and keep data in the house is far more practical than one that needs a rack of GPUs or a cloud bill. It ships with support for llama.cpp, MLX, vLLM, SGLang and ONNX from day one, and it plugs into any OpenAI-compatible backend, so trying it out is a weekend project rather than a research one.

Small, focused, honest about its limits. That's the kind of open model progress I like to see.

Sources

Reacties

Een reactie posten

Populaire posts van deze blog

SpaceX just asked to fly Starship hundreds of times a year 🚀

Last week the U.S. Federal Aviation Administration took a quiet but important step: it opened a formal review of SpaceX's request to launch Starship from Starbase in South Texas up to 456 times per year . Today the site is cleared for only 25 flights a year. That jump is the clearest signal yet that Starship is meant to become routine hardware, not a rare test rocket. What the FAA is reviewing On 14 October 2026 the FAA published a Notice of Intent to prepare a Programmatic Environmental Impact Statement (PEIS) for SpaceX's Starbase plans. A PEIS looks at a whole program of operations rather than a single launch, which is exactly what you need if you want a reusable rocket to fly the way airlines fly planes. SpaceX's proposal comes in two phases, according to the Federal Register notice: Phase 1: up to 100 launches and 200 landings per year (100 Ship, 100 Super Heavy), any day or time of week. Phase 2: up to 456 launches and 912 landings per year at Starbase. La...
Meer lezen

AI that can actually "see" a crystal 🔬

Most of the solid stuff around us, from battery materials to chips, is made of crystals: atoms stacked in neat, repeating 3D patterns. Large language models are great at reading papers about those materials, but they have never really looked at the structures themselves. A team at MIT just showed a clever way around that, and the trick is one I keep coming back to: good synthetic data. What they built The model is called CLaC, short for Contrastive Language-Crystal model, and it is described in a paper in Nature Communications by Yang Jeong Park, Mayank Kumaran, Chia-Wei Hsu, Elsa Olivetti and Ju Li. MIT News covered it on 18 October 2026. If you know CLIP, the model that taught image generators to connect pictures and words, you already get the idea. CLaC has two branches: a graph neural network that reads the 3D crystal structure, and a text encoder that reads a description. Contrastive training pulls matching structure-text pairs together in one shared space and pushes mismatche...
Meer lezen

Google's Deep Review wants to be the second pair of eyes on every long document 🔍

AI has made writing fast. Checking has not kept up. That gap is exactly what Google Research is going after with Deep Review, which went public on 14 October. It's an agentic system that reads a long technical document and tries to find what's wrong with it: broken proofs, logic gaps, unsupported claims. For anyone working on legal and document AI, like me, this is a launch worth a close look. From conference pilot to public tool Deep Review is the grown-up version of Google's Paper Assistant Tool (PAT). Over the past year, authors of more than 16,000 manuscripts at STOC 2026, ICML 2026, NeurIPS 2026 and ICLR 2027 opted in for private feedback before the submission deadline, fully separate from official peer review. The need is real: Google counts over 74,000 submissions to ICLR, NeurIPS and ICML this year, up from about 23,000 in 2023. Now anyone can use it. Every user gets two free reviews per week, and you can connect a Gemini API key for higher limits and programmati...
Meer lezen