Liquid AI launches open-weights LFM2.5-2.6B model

Liquid AI has released LFM2.5-2.6B, a text-only model with 2.69 billion parameters designed to run agentic workloads entirely on-device. Built with 30 layers—comprising 22 double-gated short convolution blocks and 8 grouped-query attention blocks—the model features a 131,072-token context window and a 128,000-token vocabulary. Pre-trained on approximately 34 trillion tokens, it is available under the lfm1.0 license in native, GGUF, MLX, and ONNX formats, supporting llama.cpp, vLLM, SGLang, and LM Studio. Two checkpoints are available: LFM2.5-2.6B-Base for fine-tuning via LoRA with TRL and Unsloth, and a post-trained version for agentic tasks.

LFM2.5-2.6B outperforms larger models on several benchmarks. On ToolSandbox, it scored 77.83, beating gemma-4-E4B-it (65.00) and Qwen3.5-9B (76.44). It also led on Multi-IF with 80.07 (beating gemma-4-E4B-it's 77.35 and Qwen3.5-9B's 62.55), IFStruct with 85.49 (versus 76.65 and 78.50), and IFBench with 59.17 (versus 39.24 and 56.47). On BFCLv4, it scored 56.88, trailing Qwen3.5-9B's 60.13 but beating gemma-4-E4B-it's 46.39. However, larger models retain an advantage in coding, with LFM2.5-2.6B scoring 59.41 on LiveCodeBenchv6 compared to Qwen3.5-9B's 69.86. Liquid AI also compared it to gemma-4-E2B-it and Qwen3.5-4B.

To build this, Liquid AI extended the vocabulary in place and used a mid-training phase to extend context to 128K across 16 languages. The base checkpoint was refined through four stages: two supervised fine-tuning rounds using an SFT mix roughly seven times larger than that of LFM2.5-8B-A1B, teacher specialization via reinforcement learning, multi-domain on-policy distillation, and agentic reinforcement learning with GRPO inside Hermes Agent and OpenClaw.

For practitioners, this enables local, cost-effective agentic deployments. The model runs at 220 tokens per second on an M5 Max in under 2.5 GB of memory, and 30 tokens per second on mobile phones. Alternatively, a single NVIDIA H100 SXM5 GPU can serve roughly 1.3 billion tokens per day. This makes it ideal for offline document triage, data extraction, and continuous background agents in regulated industries like healthcare, finance, and defense, where data privacy is paramount. However, developers should avoid using it for agentic coding or knowledge-heavy tasks.