Releases·Model Release
LiquidAI releases LFM2.5-Encoder-350M, a 15-language bidirectional encoder built on LFM2 for on-device fine-tuning
Liquid AI released LFM2.5-Encoder models in 230M/250M and 350M sizes, bidirectional masked language models built on the LFM2 architecture. They target fine-tuning into task-specific models for classification, retrieval, reranking, and semantic similarity across 15 languages, with GGUF releases and a llama.cpp PR enabling browser and CPU use.
5 engagement1 sourceSat, Oct 3, 2026, 07:00 AM
AI Overview
- LFM2.5-Encoder-350M is a masked language model with full bidirectional attention, spanning 15 languages for classification, retrieval, reranking, and similarity tasks
- The encoder has 8k context, matches or beats ModernBERT throughput with a CPU long-context edge, and runs in-browser on WebGPU
- GGUF versions ship via llama.cpp PR #29862, with fill-mask demos for 230M and 350M weights on Hugging Face
Related
Releases·Model Release
Liquid AI releases LFM2.5-Encoder family for fast long-context CPU inference
Liquid AI released LFM2.5-Encoder, a family of multilingual bidirectional encoders built on the LFM2 architecture, available in 230M and 350M parameter sizes. They are designed for fast long-context inference on CPU, claiming about 3.7x speedup over ModernBERT-base at long context. The models match or beat larger encoders on GLUE, SuperGLUE, and multilingual tasks.
Liquid AI released LFM2.5-Encoder, a family of multilingual bidirectional encoders built on the LFM2 architecture, available in 230M and 350M parameter sizes. They are designed for fast long-context inference on CPU, claiming about 3.7x speedup over ModernBERT-base at long context. The models match or beat larger encoders on GLUE, SuperGLUE, and multilingual tasks.
5 engagement2 sourcesreddit, rssTue, Jul 28, 2026, 03:01 PM
Releases·Tool Release
Liquid AI merges DSpark support for LFM2 models into llama.cpp
Liquid AI's DSpark inference optimization for LFM2 models has been merged into llama.cpp via PR #27383, enabling up to 3.2x faster inference. Official GGUF files are available on Hugging Face, and the models are small enough to run on any hardware.
Liquid AI's DSpark inference optimization for LFM2 models has been merged into llama.cpp via PR #27383, enabling up to 3.2x faster inference. Official GGUF files are available on Hugging Face, and the models are small enough to run on any hardware.
31 engagement2 sourcesredditThu, Aug 20, 2026, 05:13 PM
Releases·Model Release
Liquid AI ships LFM2.5-VL-DSpark drafter, speeding LFM2.5-VL-3B decoding up to 3.13x on-device and 2.66x on H100
Liquid AI released LFM2.5-VL-DSpark, a 280M-parameter vision drafter that accelerates the LFM2.5-VL-3B model via speculative decoding, with day-one support in llama.cpp, MLX-VLM, and SGLang. It delivers decode speedups up to 3.13x on-device and 2.66x on H100, with end-to-end gains up to 2.62x and 2.27x respectively. The drafter is open-weight and available on Hugging Face in Safetensors and GGUF formats.
Liquid AI released LFM2.5-VL-DSpark, a 280M-parameter vision drafter that accelerates the LFM2.5-VL-3B model via speculative decoding, with day-one support in llama.cpp, MLX-VLM, and SGLang. It delivers decode speedups up to 3.13x on-device and 2.66x on H100, with end-to-end gains up to 2.62x and 2.27x respectively. The drafter is open-weight and available on Hugging Face in Safetensors and GGUF formats.
0 engagement1 sourcerssThu, Sep 24, 2026, 02:08 PM
Releases·Model Releasefeatured
LiquidAI ships LFM2.5-24B-A2B, a 24B-parameter MoE reasoning model with 2.45B active parameters for on-device agents
Liquid AI released LFM2.5-24B-A2B, an open-weight 24B-parameter MoE reasoning model with ~2.45B active parameters, built for fast, private on-device agents. It scores 48.09 on LiveClaw versus 38.19 for Qwen3.6-35B-A3B despite ~50% fewer parameters, and runs 184 tok/s on an Apple M5 Max. It is the largest and final model in the LFM2.5 family, which spans 230M to 24B.
Liquid AI released LFM2.5-24B-A2B, an open-weight 24B-parameter MoE reasoning model with ~2.45B active parameters, built for fast, private on-device agents. It scores 48.09 on LiveClaw versus 38.19 for Qwen3.6-35B-A3B despite ~50% fewer parameters, and runs 184 tok/s on an Apple M5 Max. It is the largest and final model in the LFM2.5 family, which spans 230M to 24B.
726 engagement2 sourcesreddit, rssTue, Oct 13, 2026, 05:05 PM
Social·Community
Users report LFM2.5 2.6B underperforms Qwen3.5 4B on tool calling in OpenCode
A developer tested Liquid AI's LFM2.5 2.6B dense model on agentic tool calling and coding with OpenCode, finding it vastly underperformed Qwen3.5 4B on tool calling. The model struggled with actual tool calls and was confused about the working directory, despite running at ~90 t/s with Q8 on an M5 Pro.
A developer tested Liquid AI's LFM2.5 2.6B dense model on agentic tool calling and coding with OpenCode, finding it vastly underperformed Qwen3.5 4B on tool calling. The model struggled with actual tool calls and was confused about the working directory, despite running at ~90 t/s with Q8 on an M5 Pro.
24 engagement1 sourceredditWed, Aug 5, 2026, 06:15 PM