Liquid AI 发布 LFM2-2.6B 模型
Introducing LFM2-2.6B: Redefining Efficiency in Language Models | Blog | Liquid AI
Liquid AI 发布 LFM2 系列目前最大的模型 LFM2-2.6B，参数规模 2.6B，基于 10 万亿 token 训练，在 GSM8K 上达到 82.41%、IFEval 上达到 79.56%，官方称其表现超过 Llama 3.2-3B-Instruct、Gemma-3-4b-it 和 SmolLM3-3B 等 3B+ 级别模型。该模型主要针对英语和日语调优，同时保持法语、西班牙语、德语、意大利语、葡萄牙语、阿拉伯语、中文和韩语的性能。其混合架构交替使用 Grouped Query Attention（GQA）与短卷积层，官方称可带来更快的推理速度和更低的 KV cache 占用。模型已在 Hugging Face 上以 LFM Open License 开放。
推荐理由
AI 评分 52
LFM2 系列最大模型以 2.6B 参数对标 3B+ 级别，读者可据此了解小模型在推理与部署成本上的取舍。
主题
正文
当前显示原文，中文全文译文尚未就绪。
On this page
We're excited to announce LFM2-2.6B, the newest and currently largest model in our Liquid Foundation Model 2 series. Building on our 350M, 700M, and 1.2B models, LFM2-2.6B proves that exceptional performance doesn't require massive scale.
Small Model, Big Impact
With just 2.6 billion parameters, LFM2-2.6B outperforms models in the 3B+ class including Llama 3.2-3B-Instruct, Gemma-3-4b-it, and SmolLM3-3B. Trained on 10 trillion tokens, it achieves an impressive 82.41% on GSM8K for mathematical reasoning and 79.56% on IFEval for instruction following—rivaling much larger models.
The model was specifically tuned for English and Japanese, while maintaining strong performance in French, Spanish, German, Italian, Portuguese, Arabic, Chinese, and Korean, making it truly versatile for global applications.
Speed Through Innovation
LFM2-2.6B's hybrid architecture alternates blocks of Grouped Query Attention (GQA) with short convolutional layers, delivering significantly faster inference speeds and reduced KV cache requirements compared to other modern architectures. This translates to faster responses and lower deployment costs at scale.
Open and Available
LFM2-2.6B is now available on Hugging Face under our LFM Open License, making cutting-edge efficient AI accessible to developers and researchers worldwide.
The LFM2 series continues to push the boundaries of efficient AI. We're proving that with the right architecture and approach, smaller models can deliver enterprise-grade performance without the computational overhead. In the future, we will continue to scale our foundation models to bring this level of efficiency to more devices and unlock new applications.
Share on social:
For press inquiries, email us at press@liquid.ai