Liquid AI releases LFM2.5-24B-A2B, an open-weight 24B model for local agents

Liquid AI released LFM2.5-24B-A2B, an open-weight reasoning model with about 2B active parameters, built for tool-calling agents on local hardware.

Liquid AI released LFM2.5-24B-A2B on October 13, 2026. It is an open-weight mixture-of-experts reasoning model with 24B total parameters and about 2B active per token, aimed at agents that plan and call tools on your own machine. Base and post-trained weights are on Hugging Face now. For builders, it is another candidate for private agent loops that make no cloud API calls.

The headline numbers are Liquid's own. The company reports 184 tokens/s on an Apple M5 Max and 114 on an AMD Ryzen AI Max+ 395. Unite.AI notes those runs used llama.cpp at Q4_K_M with a 4K-token input. On Liquid's benchmark table, the model leads or matches Qwen3.6-35B-A3B on five of eight agentic benchmarks, including LiveClaw (48.09 vs 38.19) and AutomationBench (27.69 vs 8.72). It trails on coding and knowledge tasks, such as SWE-Bench Verified (51.00 vs 66.20). Liquid recommends pairing it with retrieval or web search. No independent reproduction has turned up yet.

It is reasoning-only, so it always writes a chain of thought before answering. Its context window is 131,072 tokens. Day-one runtimes are llama.cpp, MLX, vLLM, SGLang and ONNX. The official GGUF repo gives a starting command:

llama-cli -hf LiquidAI/LFM2.5-24B-A2B-GGUF \
  --temp 0.1 --top-k 50 --repeat-penalty 1.05

Memory is the practical limit, because all 24B weights still have to load. Enclave AI estimates you need at least a 24 GB Mac for the 13.6 GB 4-bit MLX build, since the KV cache needs memory on top. Before any commercial rollout, check the license. TechRepublic reports that it has a revenue threshold for commercial use. Hosted API access is $0.10 per 1M input tokens and $1.00 per 1M output tokens.

Sources