Liquid AI Releases LFM2.5-24B-A2B for Fast On-Device Agentic Workflows
What changed
Liquid AI has published LFM2.5-24B-A2B, an open-weight reasoning model engineered specifically for autonomous agent workflows. The architecture retains a 24-billion parameter foundation but activates only 2 billion parameters per token through a mixture-of-experts configuration. This setup utilizes top-4 routing across 64 experts within 40 layers, paired with gated short convolutions and grouped query attention. The context window expanded from 32,000 to 131,000 tokens, while an in-place tokenizer expansion doubled the vocabulary from 65,000 to 128,000 tokens to improve non-Latin language handling. Unlike earlier iterations, this release functions strictly as a reasoning model that generates an explicit chain of thought before committing to a final output.
Training followed a structured progression. Initial pre-training scaled from approximately 17 trillion to 33 trillion tokens. Mid-training then divided into two distinct phases: 2 trillion tokens processed at a 32K context length focused on reasoning, mathematics, coding, instruction compliance, and multilingual text, followed by 300 billion tokens trained at the extended 131K context length prioritizing long-document comprehension. Agentic behaviors were integrated throughout via supervised fine-tuning, teacher specialization, multi-domain on-policy distillation, and reinforcement learning conducted directly inside agent environments like OpenClaw, Hermes Agent, Pi, and OpenCode.
In direct evaluations, LFM2.5-24B-A2B competes specifically with similarly sized mixture-of-experts models, most notably Qwen3.6-35B-A3B. It outperforms its primary MoE rival on AutomationBench (27.69 versus 8.72), BFCLv4 (65.01 versus 65.00), ToolSandbox (80.71 versus 75.02), LiveClaw (48.09 versus 38.19), and τ³-Bench Banking (8.04 versus 7.63). It also leads all three instruction-following benchmarks tested. For broader context on how open models stack up in enterprise environments, you can review our analysis of Gemini 4 Argon’s engineering benchmarks.
Why it matters
Running autonomous agents locally requires balancing capability with computational overhead. By restricting active parameters to 2 billion per token, LFM2.5-24B-A2B significantly reduces memory bandwidth demands and inference latency compared to dense alternatives. This efficiency translates to measurable throughput gains: the model decodes at 184 tokens per second on an Apple M5 Max chip and reaches 114 tokens per second on AMD’s Ryzen AI Max+ 395 processor. On server-grade hardware, a single NVIDIA B300 GPU can push over 21,000 output tokens per second under high concurrency.
The explicit reasoning step means the model structures its tool-calling logic before committing to actions, which improves reliability in multi-step agent harnesses. Combined with day-one compatibility across llama.cpp, MLX, vLLM, SGLang, and ONNX, developers can deploy private, always-on agents without relying on third-party APIs or incurring unpredictable cloud costs.
What to do next
To integrate the model into your stack, start by upgrading your transformer library to version 5.0.0 or newer. Load the weights using standard PyTorch pipelines, specifying device_map="auto" and dtype="bfloat16". If your GPU supports Flash Attention 2, uncomment the corresponding implementation flag to maximize efficiency.
For agent orchestration, serve the model behind any OpenAI-compatible API endpoint and route your harness configuration accordingly. Port assignments vary by runtime: llama.cpp and MLX default to 8080, vLLM uses 8000, SGLang listens on 30000, and LM Studio operates on 1234. Set your harness to expect chat completions, enforce tool-use validation, and allocate the full 131,072 token context window. Fine-tuning guides are available for adapting the base weights to domain-specific trajectories.
Limitations & Caveats While LFM2.5-24B-A2B excels at tool invocation and procedural task completion, it trails larger peers on knowledge-dense and complex coding benchmarks. Scores on GPQA Diamond, HLE, AA-Omniscience, SWE-Bench Verified, and SciCode remain lower than comparable 30B+ models. For workloads requiring deep factual recall or extensive software engineering verification, pair this model with retrieval-augmented generation or external web search to compensate for its smaller active parameter footprint.
Sources:
Sources
Share Article
If this article helped you, please share it with others!