LFM2.5-24B-A2B: A 2B-Active Agent Model for Your Laptop

Liquid AI has released LFM2.5-24B-A2B, the largest model yet in its on-device LFM2.5 family. Announced on October 13 with open weights, it is a mixture-of-experts model that stores 24 billion parameters but uses only about 2 billion for each token. Liquid AI built it for one job: running fast, tool-calling AI agents on hardware you already own, not in the cloud.

• Sparse by design: 24B total parameters, about 2B active per token, with 64 experts per layer and 4 routed at a time
• Long context: 131K tokens, up from 32K in the previous LFM2-24B-A2B
• Fast decode: Liquid AI reports 184 tokens/s on Apple M5 Max and 114 tokens/s on Ryzen AI Max+ 395
• Day-zero tooling: support in llama.cpp, MLX, vLLM, SGLang and ONNX, with base and post-trained weights published

Why a 2B-Active Model Feels Fast Locally

Mixture-of-experts models split their knowledge across many specialist sub-networks and wake only a few for each token. LFM2.5-24B-A2B has 40 layers and 64 experts per block, with top-4 routing. It keeps the LFM2 hybrid of gated short convolutions and grouped-query attention, which keeps memory traffic low on consumer chips. Generation speed is limited mostly by the parameters touched per token, not the total. That is why a 24B model can feel as responsive as a much smaller dense one.

Liquid AI also doubled the vocabulary from 65K to 128K tokens, which usually means fewer tokens per sentence and better handling of code and non-English text. The company says the model uses the least memory of the MoE models it compared on edge hardware. On a single NVIDIA B300 at high concurrency, it reports more than 21,000 output tokens per second, so the same weights scale up to servers too.

How Does LFM2.5-24B-A2B Compare to Qwen3.6-35B-A3B?

Liquid AI measured the model against Qwen3.6-35B-A3B, a larger MoE with roughly 50% more total and active parameters. By the company's own numbers, the agent and instruction-following results stand out:

• AutomationBench: 27.69 vs 8.72
• LiveClaw: 48.09 vs 38.19
• ToolSandbox: 80.71 vs 75.02
• IFEval: 96.75 vs 93.06, and IFBench 74.69 vs 64.21

Liquid AI is candid that the larger Qwen model leads on knowledge- and coding-heavy tests. It scores higher on GPQA Diamond (83.69 vs 74.42) and SWE-Bench Verified (66.20 vs 51.00). Liquid AI recommends pairing its model with retrieval or web search for knowledge-heavy tasks. These are vendor-run benchmarks, all run on vLLM with each benchmark's recommended settings, and independent evaluations have yet to land.

Built for Agent Harnesses

The new model is reasoning-only: it always thinks before answering. It was post-trained specifically for tool calling and multi-step tasks. Liquid AI names agent harnesses such as OpenClaw, Hermes Agent, Pi and OpenCode as targets. It also publishes fine-tuning tutorials for teams that want to specialize the model.

For local AI builders, that combination is the story. A desktop agent that reads files, calls APIs and follows instructions closely now has a capable open option that runs at interactive speeds on a laptop or mini PC. Our earlier look at LFM2.5-2.6B running agents in 2.5GB of RAM showed the small end of this family. The 24B model is the step up for machines like the 192GB Ryzen AI Halo.

Who Should Try It?

If you run local models through llama.cpp or MLX and lean on tool use more than trivia, LFM2.5-24B-A2B is worth a download. Its 131K context leaves room for long agent transcripts, and its sparse design keeps responses fast without a datacenter GPU. Follow more open-model releases in our AI coverage.

Sources: Liquid AI on Hugging Face — LFM2.5-24B-A2B announcement — October 13, 2026; LLM Reference — LFM2.5-24B-A2B model profile — October 2026.