Liquid AI LFM2.5-24B-A2B: open-weight 24B agent model for on-device agents
Liquid AI released LFM2.5-24B-A2B, a 24B-parameter MoE reasoning model with 2B active per token, 131K context, day-one support in llama.cpp, MLX, vLLM, SGLang and ONNX, built for agents that run on local hardware.
Why it matters
Trained with reinforcement learning inside real agent harnesses (OpenClaw, Hermes Agent, Pi, OpenCode) rather than on static tasks; both base and post-trained weights are open, so a capable local agent is a download instead of an API bill — vendor benchmarks beat larger MoE rivals on several agentic tasks.
What you could build with it
An indie developer could ship a fully offline desktop agent that plans and calls tools on a laptop, pairing this model with a local browser-control harness to run multi-step research and filing jobs without any API spend.
Does it hold up?
Promising but unverified: vendor benchmarks beat larger rivals on some agentic tasks (AutomationBench) and trail on others (SWE-Bench Verified); the open weights mean independent tests are now possible, but none are published yet.
Built with Liquid AI LFM2.5-24B-A2B
- alfw.ai: Liquid LFM2.5-24B — a 24B agent that thinks with 2Balfw.ai · Independent writeup on the MoE trade-off pushed toward the edge: 24B stored, about 2.45B active per token, context quadrupled to 131K and vocabulary doubled to 128K for non-Latin text, with RL done inside real agent harnesses.
- GIGAZINE: LFM2.5-24B-A2B runs the Hermes Agent locally on an M5 Max MacBook Progigazine · Demo coverage: the Q4_K_M quantized model decodes at 184 tok/s on an M5 Max MacBook Pro while running the Hermes Agent on a Hugging Face briefing task; day-one llama.cpp, MLX, vLLM, SGLang and ONNX support.
Learn more
First spotted on article: source.
More AI models
OpenAI publishes 722 AI-generated math manuscripts from an unreleased modelOpenAI released 722 mathematical manuscripts in 372 result families to GitHub under Apache-2.0, produced by an…models · JEV 0.74ElevenLabs Eleven v4 + v4 Turbo: new emotive TTS architecture with a 100ms real-time variantElevenLabs released Eleven v4 and Eleven v4 Turbo on September 28: a new text-to-speech architecture with inline…models · JEV 0.72Mistral Large 4 (Le Chonk): 1.05T-param open-weight MoEMistral's new flagship: a 1.05T-parameter mixture-of-experts multimodal model with 49B active params and 1M-token…models · JEV 0.72VoiceStudio: open-source, fully-local ElevenLabs alternative with voice cloning, dubbing and dictation in 646 languagesVoiceStudio is an open-source, fully-local voice AI studio: voice cloning from a 3-second sample, voice design, video…models · JEV 0.71JetBrains Mellum 2.1: 12B MoE coding model trained for agentsJetBrains released Mellum 2.1, a 12B MoE model (2.5B active parameters, 128K context) under Apache 2.0, trained mainly…models · JEV 0.69DeepSeek V4-Flash official API: public beta with upgraded agent capabilities, Responses API and Codex supportDeepSeek's official V4-Flash is now a public beta API with massively upgraded agent capabilities - benchmark scores…models · JEV 0.69
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.