Liquid AI’s LFM2.5-24B-A2B Brings Private Agents to Your Laptop

Liquid AI dropped LFM2.5-24B-A2B on Oct. 13, 2026, pitching it as a reasoning model purpose-built to plan tasks, call tools and run autonomously — all on hardware you already own. The model is open-weight, free to download and available now on Hugging Face.

The architecture is what makes the on-device pitch credible. LFM2.5-24B-A2B holds 24 billion total parameters but activates only 2 billion of them per token, thanks to a Mixture-of-Experts design with 64 experts per block and top-4 routing. That keeps the per-token compute cost close to a 2B dense model even though the full network is much larger. On an Apple M5 Max, the model decodes at 184 tokens per second — up to 1.7 times faster than the next-best comparable model in Liquid AI’s own testing. On a single NVIDIA B300 GPU at high concurrency, throughput climbs above 21,000 output tokens per second.

Every response begins with an explicit chain of thought before a final answer, a design Liquid AI baked in from the ground up rather than bolted on. Pre-training scaled from roughly 17 trillion tokens on the predecessor model to roughly 33 trillion here, and the context window grew from 32,000 to 131,000 tokens — long enough to ingest an entire codebase before reasoning about it. The vocabulary also doubled, from 65,000 to 128,000 tokens, improving handling of non-Latin scripts.

Post-training is where the model’s agentic character takes shape. Liquid AI ran five sequential stages: supervised fine-tuning, teacher specialization, two rounds of multi-domain on-policy distillation (one non-agentic, one agentic), and finally agentic reinforcement learning conducted inside real agent harnesses — OpenClaw, Hermes Agent, Pi and OpenCode. The result is a model that has literally practiced completing multi-step tasks inside the same frameworks developers use today.

How It Stacks Up Against the Competition

Liquid AI’s primary benchmark target is Qwen3.6-35B-A3B, Alibaba’s competing local MoE, which carries roughly 50% more total and active parameters. On LiveClaw, an agentic evaluation benchmark, LFM2.5-24B-A2B scored 48.09 against Qwen3.6-35B-A3B’s 38.19. Across 16 benchmarks covering tool use, instruction following, math, knowledge and coding, LFM2.5-24B-A2B leads or matches its competitors on five of eight agentic and tool-use benchmarks and tops all three instruction-following benchmarks.

There are real limits, though. On knowledge-heavy evaluations like GPQA Diamond and HLE, and on coding benchmarks like SWE-Bench Verified, larger models maintain a clear advantage. Liquid AI openly recommends pairing this model with retrieval or web search for those workloads rather than expecting it to compete on raw knowledge recall.

One important caveat: all benchmark numbers come from Liquid AI’s own runs using vLLM. Independent third-party replication has not yet been published, so treat the figures as directionally useful rather than settled. It’s also worth noting that Qwen3.6-35B-A3B is released under the fully permissive Apache 2.0 license, while LFM2.5-24B-A2B uses Liquid AI’s own LFM Open License v1.0, which is more restrictive — a practical consideration for anyone building a commercial product on top of it.

Why This Matters If You’re a Student or Early-Career Developer

The straightforward pitch for students is cost: the model is free for non-commercial use, researchers and startups generating under $10 million in annual revenue, and it runs on hardware many people already have. A well-specced MacBook with an M5 Max chip is now sufficient to run a reasoning agent that can browse files, execute code, call external APIs and chain together multi-step workflows — entirely locally, with no data leaving the machine and no API bill accumulating in the background.

For anyone building portfolio projects in AI engineering, that combination matters. A locally running, fully observable agent stack is both a more privacy-respecting development environment and a more credible demonstration of systems-level thinking than a wrapper around a cloud API. The model ships with day-one support for llama.cpp, MLX, vLLM, SGLang and ONNX, and plugs directly into agent harnesses like Hermes Agent and OpenClaw with minimal configuration.

The 131K context window is the other practically significant number. That’s enough token space to load a substantial portion of a real codebase, a lengthy research paper or a long conversation history before asking the model to reason and act — a meaningful step toward agents that can handle realistic, not just toy, tasks.

LFM2.5-24B-A2B is the largest model in what Liquid AI describes as its final LFM2.5 family release. Both the instruction-tuned model and a base version are available for download and fine-tuning on Hugging Face today.

Source: Liquid AI via Hugging Face