Liquid AI Launches LFM2.5-2.6B: An On-Device Agentic Model With 128K Context
The new 2.69B parameter compact model features tool-calling capabilities, a 131,072-token context window, and open weights available on Hugging Face.
Stock photo for illustration only, not from the actual event
- Liquid AI releases LFM2.5-2.6B with 2.69 billion parameters
- Supports up to 131,072 tokens of context length and 128K vocabulary
- Built on a hybrid architecture combining convolution and attention blocks
- Open weights are publicly available on Hugging Face under the lfm1.0 license
Liquid AI has officially rolled out its newest artificial intelligence model, LFM2.5-2.6B, designed as a compact on-device agentic model optimized for local execution on edge devices. Both model checkpoints are publicly accessible on Hugging Face under the lfm1.0 license, shipping in native, GGUF, MLX, and ONNX formats. Day-one integration support is also provided across popular frameworks including llama.cpp, vLLM, SGLang, and LM Studio.
Technically speaking, LFM2.5-2.6B features a total of 2.69 billion parameters distributed across 30 architectural layers. Its core stack consists of 22 double-gated short convolution blocks paired with 8 grouped-query attention blocks. The model incorporates a vocabulary size of 128,000 and supports an expansive context length of up to 131,072 tokens, backed by a pre-training phase utilizing approximately 34 trillion tokens of data.
Stock photo for illustration only, not from the actual event
To achieve the expanded 128K vocabulary, Liquid AI in-place extended their existing tokenizer rather than retraining from the ground up. This was followed by a dedicated mid-training phase specifically engineered to stretch the context window to 128K tokens. The text-only model currently provides broad language coverage across 16 different languages.
The introduction of highly capable small-scale models with advanced tool-calling and long-context capabilities on edge devices highlights a major industry shift toward localized AI processing. By combining efficient convolutional layers with attention mechanisms, developers can handle extensive document analysis locally on hardware with restricted computational capacity without sacrificing core reasoning capabilities.
In benchmark evaluations against larger alternatives such as gemma-4-E2B-it (5.1B), gemma-4-E4B-it (8B), Qwen3.5-4B (4.7B), and Qwen3.5-9B (9.7B), LFM2.5-2.6B demonstrates leading performance across instruction-following benchmarks and nearly sweeps tool-use evaluations, trailing only behind Qwen3.5-9B on BFCLv4. Nonetheless, larger models maintain a traditional edge in software coding tasks, where LFM2.5-2.6B records a score of 59.41 on LiveCodeBenchv6 compared to 69.86 achieved by Qwen3.5-9B.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment