LFM 2.5-2.6B

LFM 2.5-2.6B is Liquid AI's open-weight dense model built for on-device agentic workloads. Released on August 4, 2026 under the lfm1.0 license, it combines a 128K context window, native tool calling, and a sub-2.5 GB memory footprint for agents running continuously on laptops, phones, and single-GPU servers.

Core Features

• Native tool calling: Pythonic and JSON function calling, trained in real agent harnesses like Hermes Agent and OpenClaw.
• 128K context, 16 languages: 131,072 tokens, 128,000-token vocabulary, including Chinese and Japanese.
• On-device speed: 220 tok/s on Apple M5 Max, 113 tok/s on AMD Ryzen AI Max+ 395, 30 tok/s on phones, under 2.5 GB memory.
• Server throughput: One H100 serves about 1.3 billion tokens daily at high concurrency.

Model Specifications

Pricing

LFM 2.5-2.6B is free. Open weights on Hugging Face, no paid hosted API tier; your only cost is your own hardware.

Use Cases

• Always-on local agents: Agents that plan and call tools 24/7 on laptops or phones.
• Tool use and extraction: Native function calling for structured extraction and API orchestration.
• RAG and long context: The 128K window fits large document sets.
• High-throughput serving: About 1.3B tokens daily from one H100.

Not suited to agentic coding (its weak spot) or knowledge-intensive tasks beyond 2.6B parameters.

Advantages

1. Leads its class on tool use and instruction following: Rivaling models four times its size via MOPD distillation and GRPO in real agent harnesses.
2. Runs anywhere: 30 tok/s on a phone, 220 tok/s on Apple M5 Max; one stack spans mobile, desktop, and server.
3. Deployment-ready: Day-one GGUF, MLX, ONNX builds plus llama.cpp, vLLM, SGLang, LM Studio.
4. Cheap at scale: High per-GPU throughput keeps serving costs low.

Tips

1. Use Q4KM: 1.67 GB, the recommended quality-footprint balance.
2. Pair it with a real harness: Hermes Agent, OpenClaw, or Pi match its training loops.
3. Skip it for coding agents: Choose a coding-focused model instead.

Conclusion

LFM 2.5-2.6B is a focused 2.6B agent engine: 128K context, native tool calling, and open weights that run almost anywhere. For agentic tool use, extraction, RAG, or long-context automation on local hardware, it is among the strongest in its class; for coding or deep world knowledge, look bigger.

Alternatives

• LFM 2.5 8B A1B: Liquid AI's larger hybrid MoE sibling.
• Qwen 3.8 27B: Bigger open-weight option with stronger general knowledge.
• Gemma 3: Google's open model family for on-device and single-GPU use.