Run fast, private agents on-device with LFM2.5-24B-A2B

Liquid AI has released LFM2.5-24B-A2B, an open-weight reasoning model built to run agent workloads locally rather than in the cloud. The model card and release blog describe a mixture-of-experts architecture in which 24 billion parameters are stored but only 2 billion are active for each token generated. That split is the core pitch: agents typically generate long chains of thought plus repeated tool calls, so a low active-parameter count keeps each reasoning token cheap enough to run continuously on a user's own hardware. Both LFM2.5-24B-A2B and its base counterpart, LFM2.5-24B-A2B-Base, are available on Hugging Face now.
The architecture inherits from LFM2-24B-A2B: 40 layers, 64 experts per MoE block with top-4 routing, and a hybrid of gated short convolutions and grouped-query attention. Two changes matter for agent work. The context window expands from 32K to 131K tokens, which gives an agent room for long tool transcripts and multi-turn sessions. An in-place tokenizer expansion doubles the vocabulary from 65K to 128K tokens, which Liquid AI says improves non-Latin tokenization. The new model is reasoning-only, meaning it emits an explicit chain of thought before its final answer, unlike its predecessor.
The training pipeline was scaled up substantially. Pre-training grew from roughly 17 trillion to roughly 33 trillion tokens. Mid-training adds 2 trillion tokens at 32K context weighted toward reasoning, math, code, instruction following and multilingual text, followed by 300 billion tokens at 131K context weighted toward long-context data. Tool calls and multi-turn trajectories appear in both mid-training phases. Post-training runs in five stages: supervised fine-tuning, teacher specialization, then two rounds of multi-domain on-policy distillation split between non-agentic and agentic tasks, and finally agentic reinforcement learning conducted inside agent harnesses including OpenClaw, Hermes Agent, Pi and OpenCode.
Benchmarks position the model against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning across 16 benchmarks covering tool use and agentic workflows, knowledge, instruction following, math and coding. Liquid AI notes it is the smallest model in that group yet competitive with MoEs carrying roughly 50 percent more total and active parameters. On LiveClaw it scores 48.09 against Qwen3.6-35B-A3B's 38.19. It leads or matches competitors on five of eight tool-use and agentic benchmarks, and on AutomationBench it scores more than three times the next-best model. It also leads all three instruction-following benchmarks. Larger models retain an advantage on knowledge-heavy evaluations such as GPQA Diamond, HLE and AA-Omniscience, and on coding benchmarks including SWE-Bench Verified run through OpenCode and SciCode. Liquid AI's own recommendation is to pair the model with retrieval or web search for knowledge-heavy workloads.
Inference speed is the other headline. The model decodes 184 tokens per second on an Apple M5 Max and 114 tokens per second on a Ryzen AI Max+ 395, which Liquid AI says is up to 1.7 times faster than the next-best compared model while using the least memory. On GPU, it reaches over 21K output tokens per second at high concurrency on a single NVIDIA B300. Day-one support covers llama.cpp, MLX, vLLM, SGLang and ONNX. The evaluation harness used vLLM with each benchmark's recommended generation parameters.
Getting started is straightforward for anyone already working with the Transformers stack. Installation requires transformers version 5.0.0 or later, and the model loads through AutoModelForCausalLM with a chat template applied to the input. The example generation call uses temperature 0.1, top_k 50 and a repetition penalty of 1.05. For harness use, Liquid AI recommends serving the model locally behind any OpenAI-compatible backend and pointing the harness at it. Their Hermes Agent example sets a custom provider, a local base URL, a context length of 131072, chat_completions as the API mode, and enables tool-use enforcement. Default ports differ by backend: llama.cpp and MLX use 8080, vLLM uses 8000, SGLang uses 30000, and LM Studio uses 1234.
A few practical readings stand out. The MoE design means memory footprint tracks total parameters while compute tracks active ones, so a 24B model occupies far more memory than a 2B dense model even though generation cost resembles the smaller one. That trade-off is what makes 184 tokens per second plausible on a laptop-class chip, and it also explains why Liquid AI compares against 35B-class MoEs rather than dense models of similar active size. The 131K context is a meaningful jump for agents that accumulate tool outputs, but benchmark leadership on instruction following and tool use does not transfer automatically to retrieval-heavy or coding-heavy tasks, where the source explicitly notes larger models keep an edge. Developers building local coding agents should weigh that gap before replacing a hosted model outright.
Why it matters: teams building on-device or privacy-sensitive agents gain a credible open-weight option that reasons, calls tools and runs on a single machine with day-one support across the common inference runtimes. If the reported throughput holds in real deployments, that shifts some agent workloads from API calls to local hardware, changing cost, latency and data-handling assumptions. The narrow spots are knowledge-heavy and coding tasks, where pairing with retrieval or a larger model remains the safer path.
Based on reporting from the original publisher. Visit the source for full context and later updates.