模型發布／端側 AI
Liquid AI Releases LFM2.5-2.6B Weights, Bringing 128K Context and Tool Calling to a 2.69B On-Device Model
LFM2.5-2.6B uses a hybrid architecture combining short convolutions with GQA to bring agentic reasoning, tool use, and long-context capabilities to smartphones and personal computers. Liquid AI reports decoding speeds of 220 tokens/s on an M5 Max, though the performance and agent benchmarks are based primarily on the vendor’s own evaluations.
Liquid AI has released the base and agent-post-trained weights for LFM2.5-2.6B. Its goal is not to replace frontier cloud models with a small model, but to enable continuously running, privacy-sensitive, or offline agent workflows to operate locally. The model has 2.69B parameters across 30 layers, including 22 layers using double-gated short convolutions and eight using grouped-query attention. Its vocabulary has been expanded to 128,000 tokens, it supports 16 languages including Chinese, and mid-training extended its context window to 131,072 tokens. It was pretrained on approximately 34 trillion tokens.
Post-training comprised two rounds of supervised fine-tuning, domain-specific teacher specialization, multi-domain on-policy distillation, and agentic reinforcement learning. In the final stage, the model learned multi-turn tool use directly within agent harnesses such as Hermes Agent and OpenClaw, using GRPO, programmatic checks, LLM-as-a-judge evaluation, and safety gates. This more closely resembles real execution environments than training solely on fixed JSON function calls, but it may also make the results dependent on prompts and tool formats encountered during training.
Official BF16, GGUF, ONNX, and MLX variants are available, with support for Transformers, llama.cpp, vLLM, and SGLang. According to vendor testing, the model uses less than 2.5GB of memory and reaches 220 tokens/s on an M5 Max and 113 tokens/s on a Ryzen AI Max+ 395, while smartphones achieve approximately 30 tokens/s. Under high concurrency, a single H100 delivers close to 15,000 output tokens per second. The model scores 56.88 on BFCL v4 and 77.83 on ToolSandbox; the latter is slightly higher than the tested Qwen3.5-9B, although it still trails larger models by a wider margin on coding benchmarks.
Engineering teams should first retest time to first token, long-context KV cache behavior, tool-use accuracy after quantization, and recovery from multi-turn failures on their target hardware. Liquid AI also explicitly advises against using the model for knowledge-intensive tasks or agentic software development. In addition, the weights are distributed under the LFM Open License 1.0, so teams must review its terms before deployment rather than relying solely on “open-weight” marketing to determine usage restrictions.