AI BriefingLiquidAI Releases LFM2.5-Encoder 350M and 230M Multilingual Models
Key point
LiquidAI has released LFM2.5-Encoder-350M and LFM2.5-Encoder-230M, with the 350M variant described as a multilingual bidirectional encoder supporting 15 languages and on-device inference.
Details
LiquidAI has released LFM2.5-Encoder-350M and LFM2.5-Encoder-230M. The LFM2.5-Encoder-350M is described as a multilingual bidirectional encoder built on the LFM2 architecture, designed for on-device deployment and downstream tasks such as fine-tuning for NLI, paraphrase detection, sentiment analysis, and retrieval.
Performance and Capabilities (LFM2.5-Encoder-350M)
- Multilingual Support: The 350M model is trained across 15 languages.
- Context Window: Supports up to 8k tokens.
- Throughput: Matches or exceeds ModernBERT performance, with a noted advantage in long-context processing on CPU.
- On-Device Execution: Optimized for local inference, including support for WebGPU in browsers.
Availability
Both models are available in GGUF format for integration with llama.cpp (via PR #29862) and standard Hugging Face repositories. The release includes specific weights for the 350M and 230M parameter variants.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Unsloth Dynamic 3.0 GGUF Released
OpenDLSS-NR: Bit-Exact Vulkan Reimplementation of NVIDIA DLSS 5 Neural Rendering Network