LFM2.5-24B-A2B: Inside LiquidAI's New Agentic MoE Model

LiquidAI released LFM2.5-24B-A2B, a 24-billion-parameter mixture-of-experts model that activates about 2.45 billion parameters per token and supports a 131,072-token context window, the company's largest and final release in the LFM2.5 family. LiquidAI reports the model scores 48.09 on the LiveClaw agentic/tool-use benchmark versus 38.19 for Qwen3.6-35B-A3B, which has roughly 50% more total and active parameters, and quotes 184 tokens/sec on an Apple M5 Max and up to 21K output tokens/sec on a single NVIDIA B300 GPU. The model ships day one in native safetensors, FP8, GGUF, ONNX and MLX (4-bit) formats, and LiquidAI recommends pairing it with retrieval or web search rather than relying on it for knowledge-heavy tasks.

LFM2.5-24B-A2B: Inside LiquidAI's New Agentic MoE Model LiquidAI's LFM2.5-24B-A2B packs 24B total/2B active params, 131K context, and strong agentic benchmarks for on-device reasoning. What is LFM2.5-24B-A2B? LFM2.5-24B-A2B is LiquidAI’s largest and final release in the LFM2.5 model family: a 24-billion-parameter mixture-of-experts MoE model that only activates about 2.45 billion parameters per token. It’s built specifically for agentic use cases: tool calling, multi-step reasoning, and running on local hardware rather than in a cloud datacenter. LiquidAI positions it as a reasoning model that always thinks before answering, competing against MoE models considerably larger than itself on agentic benchmarks. TL;DR - Sparse by design : the model has 24B total parameters but only activates 2.45B per token, using 64 experts per MoE block with top-4 routing across 40 layers. - Agentic benchmark claim : LiquidAI reports a LiveClaw score of 48.09, ahead of Qwen3.6-35B-A3B’s 38.19, despite that model having roughly 50% more total and active parameters. - Long context : the model supports a 131,072-token context window, up from the 32K window on the prior LFM2-24B-A2B generation. - Built for edge hardware : LiquidAI quotes 184 tokens/sec on an Apple M5 Max and up to 21K output tokens/sec on a single NVIDIA B300 GPU, with up to 1.7x faster inference than comparable MoEs on-device. - Trained inside real agent harnesses : post-training reinforcement learning happened inside OpenClaw, Hermes Agent, Pi, and OpenCode, so the model has learned the tool-calling conventions those frameworks actually use. - Multiple deployment formats ship day one : native safetensors, FP8, GGUF, ONNX, and MLX 4-bit versions are all available, covering GPU servers, CPUs, and Apple Silicon. - Not a knowledge base : LiquidAI explicitly recommends pairing it with retrieval or web search rather than relying on it for knowledge-heavy tasks on its own. How is the architecture structured? LFM2.5-24B-A2B keeps the hybrid architecture introduced in the earlier LFM2-24B-A2B model. It has 40 total layers split between 30 double-gated short convolution blocks and 10 grouped-query attention GQA blocks. The MoE layers route each token to 4 of 64 available experts. This combination of convolutional blocks and sparse attention is part of why LiquidAI can claim faster inference than comparably sized dense or MoE alternatives: most of the computation happens through lightweight paths, and only a fraction of the total parameter count gets touched per forward pass. Two concrete upgrades separate this model from its LFM2 predecessor. The context window grew from 32,000 to 131,072 tokens. And the vocabulary doubled, from 65,000 to 128,000 tokens, through what LiquidAI calls an “in-place tokenizer expansion,” aimed at improving tokenization efficiency for non-Latin scripts. The model supports 16 languages including English, Arabic, Chinese, French, German, Japanese, Korean, Spanish, Hindi, and Russian. What do the benchmark numbers actually show? The headline comparison LiquidAI makes is against Qwen3.6-35B-A3B, a model with roughly 50% more total and active parameters. On LiveClaw, an agentic/tool-use benchmark, LFM2.5-24B-A2B scores 48.09 versus Qwen3.6-35B-A3B’s 38.19. That gap is the core argument for the model: fewer active parameters, less compute per token, but competitive or better performance on the tasks it’s designed for, specifically tool use and agent workflows rather than general knowledge recall. It’s worth being precise about what this benchmark does and doesn’t tell you. LiveClaw measures agentic and tool-calling competence, not broad world knowledge or raw reasoning depth across arbitrary domains. LiquidAI itself frames the model as not ideal for knowledge-heavy tasks run in isolation, recommending it be paired with retrieval-augmented generation or live web search. That’s a meaningful caveat: this is a model optimized for a specific workload agents that need to plan, call tools, and act reliably rather than a general-purpose chatbot competing on trivia or long-form knowledge benchmarks. How fast is it in practice? Inference speed is central to the pitch. LiquidAI reports 184 tokens per second on an Apple M5 Max, and up to 21,000 output tokens per second on a single NVIDIA B300 GPU. The company also claims up to 1.7x faster inference than comparable MoE models when run on-device. Because only 2.45B parameters activate per token, the computational cost of generating each token including the reasoning tokens the model always produces before answering stays low even though the full model occupies the memory footprint of a 24B-parameter network. This matters specifically for the “thinking” behavior built into the model. Unlike the earlier LFM2-24B-A2B, this is a pure reasoning model: it always generates a <think block before producing a final answer, baked directly into its chat template. Reasoning models typically burn through more tokens per response than non-reasoning models, so keeping the per-token compute cost low is what makes that tradeoff practical on consumer or edge hardware rather than exclusively in the cloud. How was it trained? Other agents ship a demo. Remy ships an app. Real backend. Real database. Real auth. Real plumbing. Remy has it all. LiquidAI describes a training pipeline that scaled substantially beyond the prior generation. Pre-training used roughly 33 trillion tokens, nearly double the ~17 trillion used for LFM2. Mid-training ran in two phases: 2 trillion tokens at a 32K context window on curated data weighted toward reasoning, math, code, instruction following, and multilingual text, followed by 300 billion tokens of long-context data extending the window to 131K. Tool-call examples and multi-turn trajectories were included in both mid-training phases, not bolted on afterward. Post-training involved supervised fine-tuning, teacher specialization, two rounds of multi-domain on-policy distillation one non-agentic, one agentic , and agentic reinforcement learning conducted inside actual agent harnesses: OpenClaw, Hermes Agent, Pi, and OpenCode. Training the model directly within these frameworks means it picks up the specific system prompts, tool schemas, and interaction patterns those harnesses use in production, rather than learning a generic approximation of tool-calling behavior. Is LFM2.5-24B-A2B worth deploying? For teams building agents that need to run locally, call tools, and keep latency and cost predictable, it’s a reasonable candidate to evaluate. The combination of a 2B active-parameter footprint with 24B total capacity targets a specific gap: enough capacity to handle varied tool-calling scenarios, but cheap enough per-token to run on a laptop-class GPU or a single high-end data center card without the throughput penalty of a dense model that size. The tradeoffs are clear from LiquidAI’s own framing. It’s not meant to be a standalone knowledge engine, so if your use case depends heavily on recalling facts rather than orchestrating tools, you’ll want retrieval or search bolted on regardless. It’s also a pure reasoning model, meaning every response involves a thinking phase, which adds latency and token cost compared to non-reasoning models for simple queries. And benchmark comparisons like the LiveClaw score, while favorable, come from a single agentic benchmark rather than a broad suite, so it’s worth testing against your own tool-use scenarios before committing. Deployment flexibility is a real strength here. LiquidAI shipped the model simultaneously in native safetensors format, FP8 for high-throughput GPU serving with lower memory , GGUF for llama.cpp and CPU inference , ONNX for cross-platform deployment , and MLX 4-bit for Apple Silicon . A speculative decoding draft model DSpark is listed as coming soon for SGLang. That range of formats, available from day one on Hugging Face, lowers the integration cost for teams already using vLLM, SGLang, llama.cpp, or LM Studio. Frequently Asked Questions How many parameters does LFM2.5-24B-A2B actually use at inference time? The model has 24 billion total parameters but activates only about 2.45 billion per token, thanks to its mixture-of-experts design with 64 experts per MoE block and top-4 routing. What is LiveClaw and why does the benchmark comparison matter? LiveClaw is an agentic and tool-use benchmark. LiquidAI reports LFM2.5-24B-A2B scoring 48.09 versus 38.19 for Qwen3.6-35B-A3B, a model with roughly 50% more total and active parameters, making the comparison a claim about efficiency per parameter on agent-style tasks specifically. Does LFM2.5-24B-A2B support long documents or long conversations? Yes. It has a 131,072-token context window, up from the 32K window in the earlier LFM2-24B-A2B model, enabled by additional long-context mid-training on 300 billion tokens. Can I run this model on a Mac or consumer hardware? Yes. LiquidAI provides an MLX 4-bit build for Apple Silicon and reports 184 tokens per second on an Apple M5 Max, along with GGUF builds for CPU inference via llama.cpp and LM Studio. Is this a good model for general knowledge or chatbot use? Remy doesn't build the plumbing. It inherits it. Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something. Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want. LiquidAI specifically recommends against relying on it for knowledge-heavy tasks alone, suggesting it be paired with retrieval or web search. It’s optimized for agentic reasoning and tool use rather than acting as a standalone knowledge base.