The development

What happened

Liquid AI released LFM2.5-24B-A2B on October 13, pairing a mixture-of-experts architecture with reasoning and tool-use training for local agents. The model has 24 billion total parameters but activates two billion per token. Its context window grows to roughly 131,000 tokens, and both reasoning and base weights are available through Hugging Face.

Key facts

  • The release extends the predecessor’s 32K-token context window to about 131K tokens and supports runtimes including llama.cpp, MLX, vLLM, SGLang and ONNX.

  • Liquid AI reports 184 output tokens per second on an Apple M5 Max. Its separate result above 21,000 tokens per second on an NVIDIA B300 uses high concurrency, not one user's generation speed.

  • The company’s table reports 51.00 on SWE-Bench Verified versus 66.20 for Qwen3.6-35B-A3B. The smaller active-parameter count does not mean the model leads every task.

  • Two billion active parameters reduce per-token computation, but deployment still needs memory for the full weights and context.

  • Open-weight availability does not by itself establish unrestricted licensing. Local inference also does not prevent attached tools or external services from transmitting information.

The context

Why it matters

TideReach assessment: This release widens the options for agents that must run close to local data or without continuous cloud inference. Practical fit depends on total memory use, response latency and task accuracy, rather than the active-parameter headline alone. Teams can test whether the model’s tool-use strengths offset its weaker coding or knowledge results for their workloads, and examine network-connected tools separately when evaluating privacy.