Liquid AI Releases LFM2.5-24B-A2B: Open-Weight MoE Agent Model Running Locally on Laptops

On October 13, 2026, Liquid AI officially launched LFM2.5-24B-A2B, an open-weight model engineered to run autonomous AI agent workflows locally on standard consumer laptops and edge hardware. Adopting a sparse Mixture-of-Experts (MoE) architecture, the model contains 24 billion (24B) total parameters while activating only about 2 billion (2B) parameters per input token, achieving rapid local inference and robust tool execution under constrained hardware footprints.

Image source: Startup Fortune / Liquid AI

LFM2.5-24B-A2B dramatically expands upon its predecessor, LFM2-24B-A2B, extending the context window from 32K tokens to 131,072 (131K) tokens and widening the tokenizer vocabulary to 128,000 tokens. This expansion enables local agents to ingest extensive codebases and retain lengthy multi-turn execution histories directly in device memory without aggressive truncation.

24B Total Parameters with 2B Active: 4 of 64 Experts Routing and 131K Context Window

The core engineering milestone of LFM2.5-24B-A2B lies in its sparse MoE routing scheme, which preserves the representational capacity of a large model while keeping compute requirements down to on-device levels.

Each MoE block contains 64 specialized expert networks, dynamically routing each incoming token to the top 4 most relevant experts during inference.

• 24B Total Scale with ~2B Active Load: While housing 24B parameters in memory, the model expends only ~2B compute per token, preventing thermal throttling and excessive battery draw during continuous local operation.
• Top-4 Routing per MoE Block: Selecting 4 of 64 experts provides the necessary behavioral flexibility to coordinate complex function-calling sequences and multi-step tool interactions.
• 131K Context Window: Expanding to 131,072 tokens represents a fourfold increase over the previous 32K baseline, allowing developers to maintain multi-file code context and agent tool logs in memory.
• 128,000 Token Vocabulary: An expanded tokenizer vocabulary enhances encoding efficiency across programming languages and multilingual inputs.

Local M5 Max and Ryzen Inference Benchmarks Alongside Agent Tool Performance

Hardware measurements and standardized benchmarks demonstrate notable execution efficiency across modern client platforms.

In local inference benchmarks utilizing llama.cpp with Q4_K_M 4-bit quantization, LFM2.5-24B-A2B achieved 184 tokens per second on an Apple M5 Max MacBook and 114 tokens per second on an AMD Ryzen AI Max+ 395 processor. In high-concurrency server deployments, a single Nvidia B300 GPU achieved peak throughput of up to 21,000 output tokens per second.

The model also demonstrated strong results across agent tool execution and instruction-following benchmarks compared to competing open-weight architectures:

• LiveClaw (Agent Tool Use): Scored 48.09, outperforming Qwen3.6-35B-A3B (38.19) by nearly 10 points.
• ToolSandbox (Tool Execution): Reached 80.71, surpassing Qwen3.6-35B-A3B's 75.02.
• IFEval (Instruction Following): Attained a 96.75 score, indicating reliable adherence to strict system prompts and output formatting constraints.

These empirical results illustrate that autonomous tool calling and structured execution loops no longer strictly require cloud-hosted frontier models to deliver responsive performance.

Open License Distribution, API Pricing, and Knowledge-Heavy Task Caveats

To encourage broad adoption across the developer community, Liquid AI has released LFM2.5-24B-A2B across multiple formats and permissive licensing terms.

The model weights are distributed under the LFM Open License v1.0 and hosted on Hugging Face in standard PyTorch weights, GGUF formats for llama.cpp, and Apple silicon-optimized MLX formats. For organizations preferring managed cloud infrastructure, Liquid AI also provides an official hosted API priced at $0.10 per million input tokens and $1.00 per million output tokens.

However, teams evaluating the model should account for verified architectural trade-offs and operational caveats:

• External Retrieval (RAG) Recommended for Knowledge Tasks: Because the model activates only ~2B parameters per token, it is not recommended for standalone deployment on knowledge-heavy factual queries. Reliable enterprise deployments should pair the model with live search tools or internal RAG retrieval pipelines.
• Slight Variance in Raw Math Reasoning: On the MATH-500 benchmark, LFM2.5-24B-A2B posted a score of 96.04, slightly trailing Qwen3.6-35B-A3B (96.73).

Liquid AI's latest release reflects the rapid maturation of on-device agent architectures, demonstrating that high-capacity sparse routing and extended context windows can deliver practical, privacy-preserving agent automation directly on client hardware.

Sources

• Startup Fortune: Liquid AI's LFM2.5-24B-A2B Runs Agent Tasks Locally on a Laptop