Liquid AI ships LFM2.5-24B-A2B, its biggest on-device agent model
A 24B mixture-of-experts model that uses just 2B active parameters per token
The facts
- 24B total parameters with only 2.45B active per token, via 64 experts per MoE block and top-4 routing
- Liquid AI reports it beats Qwen3.6-35B-A3B on LiveClaw agentic benchmark, 48.09 vs 38.19, despite fewer parameters
- Context window extended to 131,072 tokens and vocabulary doubled to 128,000 tokens via in-place tokenizer expansion
- Company claims up to 1.7x faster on-device inference than comparable MoEs: 184 tok/s on Apple M5 Max, 21K tok/s on Nvidia B300
Liquid AI has released LFM2.5-24B-A2B, a 24-billion-parameter reasoning model. It activates only 2.45 billion parameters per token, thanks to a mixture-of-experts design with 64 experts per block and top-4 routing. It is the largest and final entry in the company's LFM2.5 family. The family now runs from the 230-million-parameter LFM2.5-230M up to this model, spanning two orders of magnitude on a single architecture lineage built for on-device agents.
Liquid AI says the model is competitive with Qwen3.6-35B-A3B on agentic and tool-use tasks, despite that rival having roughly 50% more total and active parameters. On the LiveClaw benchmark it scores 48.09, against 38.19 for the larger model, the company reports. Liquid AI also says the model runs up to 1.7 times faster than comparable MoE models on-device. It hits 184 tokens per second on an Apple M5 Max, and as much as 21,000 output tokens per second on a single Nvidia B300 GPU. Context length grows to 131,072 tokens from 32,000 in the prior LFM2-24B-A2B. The vocabulary doubles to 128,000 tokens via an in-place tokenizer expansion, aimed at better non-Latin language handling. Pre-training used about 33 trillion tokens, up from roughly 17 trillion. That was followed by mid-training on curated reasoning, math, code and multilingual data, then post-training that includes on-policy distillation and agentic reinforcement learning inside harnesses such as OpenClaw, Hermes Agent, Pi and OpenCode.
Unlike its predecessor, LFM2.5-24B-A2B always reasons before answering. It inserts a think tag directly into its chat template, rather than leaving reasoning optional. Liquid AI ships the model in several formats: an FP8 checkpoint for GPU serving with vLLM and SGLang, a GGUF build for llama.cpp, an ONNX version for cross-platform deployment, and an MLX 4-bit build for Apple Silicon. A draft model for speculative decoding is listed as coming soon. The company recommends pairing it with retrieval or web search for knowledge-heavy tasks, since it is built chiefly for reasoning and tool use rather than as a standalone knowledge store.
Why it matters
For teams building local or edge agents, a 2B-active-parameter model matching a 35B-active rival on tool use changes the cost-per-agent-call math. Its reasoning-only chat template and quantised formats make it a plausible drop-in for existing agent harnesses, not a research curiosity.
Written from the primary source above by The Neural Net’s automated desk (claude-sonnet-5, theneuralnet/pipeline@1) without human review. How this works. Corrections: check the source first — if we got it wrong, the source link is the record.