Liquid AI releases LFM2.5 encoders for long-context tasks on CPU
tomaarsen · x · 2026-07-29
Liquid AI’s LFM2.5-Encoder-230M and LFM2.5-Encoder-350M are now on Hugging Face. The company says the bidirectional encoders are derived from its LFM2.5 backbones by switching from causal to bidirectional attention, making short convolutions non-causal, and training with a masked language modeling objective. The models target classification, natural language understanding, token-level tasks, and retrieval, and are designed to stay fast on long inputs, including CPU-only setups.
Related event: LiquidAI Releases Two Multilingual Encoders Optimized for CPU(3 posts)→
More from Models
- Together AI and Moonshot AI set July 30 webinar on Kimi K3 architecture — togethercompute · 2026-07-29
- Three frontier models missed a simple link-extraction task and rewrote it as a web app — NickPassig · 2026-07-29
- Fable is being praised for trying parallel and multi-stream solutions first — dejavucoder · 2026-07-29
- Kimi K3’s intro feels unusually information-dense in very few tokens — nathanbenaich · 2026-07-29
- MoonshotAI’s FlashKDA commit hints at a hybrid attention mainline model — peterjliu · 2026-07-29
- User says Opus 5 ignores tools, breaks workflows, and feels overly ignorant — omarsar0 · 2026-07-29