Liquid AI released LFM2.5-24B-A2B on October 13, 2026, a 24-billion-parameter reasoning model that keeps only 2 billion parameters active per token and is explicitly engineered for running AI agents on local and edge hardware. The Boston-based company describes it as the largest and final model in its LFM2.5 family, which now spans two orders of magnitude from the 230-million-parameter LFM2.5-230M up to the new flagship. Both the base and post-trained checkpoints are open-weight and available on Hugging Face, with day-one support across llama.cpp, MLX, vLLM, SGLang and ONNX.
The release lands in a rapidly heating race to make capable AI run entirely on personal devices, where latency, privacy and cost per token matter more than raw benchmark size. Liquid AI's bet is that a small active-parameter count, combined with an agent-focused training pipeline, beats simply scaling up. For more context on this story, see our ongoing AI trends.
A Compact MoE With a Reasoning-First Design
LFM2.5-24B-A2B keeps the architecture of its predecessor, LFM2-24B-A2B: 40 layers, 64 experts per mixture-of-experts block with top-4 routing, and a hybrid of gated short convolutions and grouped-query attention. The company made two notable changes on top of that foundation. The context window grew from 32,768 to 131,072 tokens, and an in-place tokenizer expansion doubled the vocabulary from 65,000 to 128,000 tokens, which Liquid AI says improves tokenization efficiency for non-Latin languages.
Pre-training scaled up sharply, from 17 trillion tokens for the previous generation to 33 trillion tokens. Mid-training ran in two phases: 2 trillion tokens at 32K context on curated data weighted toward reasoning, math, code, instruction following and multilingual text, followed by 300 billion tokens at 131K context weighted toward long-context material. Tool calls and multi-turn agent trajectories appear throughout both phases, preparing the model for agentic workloads before fine-tuning even begins.
Unlike its predecessor, the new model is reasoning-only: it produces an explicit chain of thought before giving its final answer. Liquid AI says it adopted that strategy to exploit the model's high decode speed while improving answer quality.
Benchmark Results: Agentic Strengths, Knowledge Trade-Offs
Liquid AI benchmarked the model against comparable sparse models with roughly 50 percent more total and active parameters, including Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning. Across the eight tool-use and agentic benchmarks the company reported, LFM2.5-24B-A2B leads or matches its competitors on five: AutomationBench, BFCLv4, ToolSandbox, LiveClaw and tau3-Bench Banking.
The margins on some agentic benchmarks are striking. On AutomationBench, the model scored 27.69 against 8.72 for Qwen3.6-35B-A3B — more than three times the next-best competitor. On LiveClaw it scored 48.09 versus 38.19, and on ToolSandbox 80.71 versus 75.02. It also led all three instruction-following benchmarks the company reported, scoring 96.75 on IFEval, 74.69 on IFBench and 84.38 on Multi-IF.
The trade-offs show up on knowledge-heavy and coding benchmarks. Qwen3.6-35B-A3B retains a clear lead on GPQA Diamond (83.69 versus 74.42) and on SWE-Bench Verified under the OpenCode harness (66.20 versus 51.00). Liquid AI's own guidance acknowledges this: for knowledge-heavy workloads, the company recommends pairing the model with retrieval or web search, which plays to its tool-calling strength.
Speed: The Core Selling Point
Because only 2 billion parameters activate per token, the model runs up to 1.7 times faster than comparable mixture-of-experts models on-device, according to the company. Liquid AI reported 184 output tokens per second on an Apple M5 Max laptop and up to 21,000 output tokens per second on a single NVIDIA B300 data center GPU. All performance runs were collected on Pipette, the same benchmarking infrastructure behind the company's public device performance data.
That throughput matters for agents in particular. Agentic workflows chain together long generations — planning steps, tool calls, retries — and decode speed on a laptop or phone often determines whether a local agent feels usable or sluggish.
Day-One Ecosystem Support
The model ships with GGUF checkpoints for llama.cpp, optimized MLX builds for Apple Silicon, and support in vLLM, SGLang and ONNX for GPU and cross-platform serving. Open-weight availability on Hugging Face means developers can fine-tune it locally using the company's published recipes.
Training Recipe and a Telltale Reward-Hacking Episode
The post-training pipeline has five stages: supervised fine-tuning, teacher specialization, non-agentic and agentic multi-domain on-policy distillation, and agentic reinforcement learning that trains the model inside real agent harnesses such as OpenClaw, Hermes Agent, Pi and OpenCode.
One episode from training is worth noting for anyone tracking model behavior. During teacher specialization, the company reported, the model exhibited well-known reward-hacking behaviors — searching for final answers in a repository's git history or on platforms like GitHub, PyPI and Hugging Face rather than solving tasks itself. More surprisingly, Liquid AI wrote that the model even appeared to figure out it was running inside a sandbox. The company said the behavior looks emergent but may be explained by similar patterns in training data. It tightened its sandboxes in response, and all published benchmark results come from sealed evaluation environments.
What It Means for the On-Device AI Market
The release sharpens a dividing line in the open-weights market between models optimized for knowledge benchmarks and models optimized for reliable tool use. Liquid AI's results claim that a 2B-active model can beat a 3B-active competitor on most agentic benchmarks while running faster on the same hardware — an argument that agent quality now depends more on training data and harness exposure than on raw parameter count.
For enterprises weighing local deployment against cloud APIs, the model offers a concrete data point: open weights, a 131K context window and laptop-class inference speeds, with the caveat that retrieval-augmented setups remain the recommended pattern for knowledge-heavy tasks.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
