Liquid AI Releases LFM2.5-8B: 8-Billion Parameter Hybrid Model Optimized for Edge Deployment
Liquid AI has released LFM2.5-8B-A1B, an 8-billion parameter hybrid model designed specifically for edge AI and on-device deployment. The model is available in multiple GGUF quantized formats ranging from 4-bit (4.84 GB) to 16-bit (16.9 GB), optimized for memory efficiency.
Liquid AI Releases LFM2.5-8B: 8-Billion Parameter Hybrid Model Optimized for Edge Deployment
Liquid AI has released LFM2.5-8B-A1B, an 8-billion parameter hybrid model designed specifically for edge AI and on-device deployment. The company claims the model "sets a new standard in terms of quality, speed, and memory efficiency."
Technical Specifications
The model features an LFM2MoE architecture and is available in multiple GGUF quantized formats:
- 4-bit Q4_0: 4.84 GB
- 4-bit Q4_K_M: 5.16 GB
- 5-bit Q5_K_M: 6.03 GB
- 6-bit Q6_K: 6.96 GB
- 8-bit Q8_0: 9.01 GB
- 16-bit BF16/F16: 16.9 GB
The quantization range allows developers to balance model size against performance based on deployment constraints. The smallest 4-bit version requires under 5 GB of storage, making it viable for mobile and edge devices.
Deployment and Availability
The model runs via llama.cpp, the widely-used inference framework for quantized language models. According to Hugging Face data, the model has been downloaded 42 times in its first month of availability.
LFM2.5-8B-A1B is a fine-tuned version built on top of LiquidAI/LFM2.5-8B-A1B-Base. The model is part of a collection of 33 post-trained and base LFM2.5 models released by Liquid AI.
No pricing information has been disclosed for API access. The model is not currently deployed by any inference provider on Hugging Face's platform.
Architecture Details
The LFM2MoE architecture suggests a mixture-of-experts approach, though specific architectural details beyond the 8-billion parameter count have not been published. The "2" in LFM2 indicates this is the second generation of Liquid AI's hybrid models.
What This Means
Liquid AI is targeting the growing edge AI market with a model sized between typical small models (1-3B parameters) and larger general-purpose models (70B+). The 8B parameter count and aggressive quantization options suggest the company is prioritizing deployment flexibility over raw capability. However, without published benchmark scores or detailed performance comparisons, it's unclear how LFM2.5-8B compares to established edge-optimized models like Llama 3.2 3B or Phi-3.5-mini. The availability of GGUF formats makes the model immediately compatible with the broader llama.cpp ecosystem.
Related Articles
OrcaRouter releases OrcaCyber Nano 1.0, a 27B local security model with 262K context and GGUF builds
OrcaRouter has published OrcaCyber Nano 1.0, a 27B vision-language reasoning model for defensive security, as GGUF files for local inference with llama.cpp. It supports a 262,144-token context window and text, image and video input. Benchmark scores are self-reported, and the harness details are not yet published.
NVIDIA releases Nemotron 3.5 Super VL NVFP4: 120B-parameter, 12B-active vision-language model with 1M context
NVIDIA has published Nemotron 3.5 Super VL in NVFP4 quantization on Hugging Face: a 120B-parameter, 12B-active hybrid Mamba-2/Transformer MoE with a 1M-token context window and text, image and video input. NVIDIA reports the quantized checkpoint stays close to BF16 on most benchmarks. It fits on one DGX Spark, one H200, two H100s or one GB200.
Recursive releases Flint, a 40B-parameter open coding model that runs on laptops under Apache 2.0
Recursive released Flint, a free, open-weight mixture-of-experts coding model with 40 billion total and 2.5 billion active parameters, under Apache 2.0. The company claims its Eureka Machine automated research system contributed to nearly every stage of development, and that Flint runs at about 131 tokens per second on an M4 Pro MacBook Pro.
NVIDIA Kumo Relational predicts from databases using up to 10,000 in-context examples, no training
NVIDIA released Kumo Relational, a pre-trained foundation model that predicts outcomes such as churn, demand, and fraud directly on relational databases in a single forward pass. NVIDIA claims it beats supervised relational deep learning and expert-engineered gradient-boosted trees across 41 tasks on 15 databases, and surpasses AutoGluon ensembles by 8% MRR on SAP SALT.
Comments
Loading...