Across the library:
Start with the model family that matches your input and output shape, then choose a runtime based on where you want to run it. Use the complete matrix below when you need exact repository and format availability.
- Context length depends on the model: 32K tokens for most text, vision, audio, and omni models, 256K for LFM3-1.7B, 128K for LFM3-VL-2B, LFM2.5-8B-A1B, d1-2B, and d1-omni-2B, 131K for LFM2.5-24B-A2B, 8K for the LFM2.5 encoders, and 16K for d1-omni-600M. Each model page lists its exact limit.
- Runtime support varies by checkpoint and modality. Start with the model pageβs supported Transformers, llama.cpp, vLLM, SGLang, zml/llmd, MLX, or app-based path. A downloadable format does not establish every runtimeβs support for it.
- Training is also model-specific. Start with TRL for LFM3 text LoRA; use Liquid Finetune, Unsloth, or Halo for their documented checkpoints and workflows. The Trainable? column below links the modelβs path.
Model Families
Text Models
Chat, tool calling, structured output, and classification.
Vision Models
Image and video understanding, OCR, and visual reasoning.
Audio Models
Interleaved audio/text models for TTS, ASR, and voice chat.
Omni Models
Text, image, video, and audio input with text or speech output in one model.
Decision Models
Purpose-built models for classification, routing, and scoring.
Liquid Nanos
Task-specific models for extraction, summarization, retrieval for RAG (embeddings and reranking), and translation.
Common Workflows
GPU Serving
Use vLLM or SGLang for high-throughput serving, and Transformers for direct Python inference.
Local and On-Device
Use llama.cpp, Ollama, Atomic Chat, or MLX depending on platform and packaging needs. To embed a model in an iOS, Android, or desktop app, see Build with llama.cpp.
Fine-Tuning
Start with the LFM3 text LoRA smoke test in TRL. Follow Liquid Finetune, Unsloth, or Halo for supported managed, optimized, or distributed workflows.
Model Repositories
Browse LiquidAI collections on Hugging Face for model weights, GGUF exports, MLX packages, ONNX exports, and model cards.
Formats
Use the format that matches your runtime and deployment target. The matrix below links the exact repository for each model and format.- GGUF β Best for local CPU/GPU inference on any platform. Use with llama.cpp, LM Studio, Ollama, or Atomic Chat. Every open-weight model in the matrix has a
LiquidAI/<model>-GGUFrepository. - MLX β Best for Mac users with Apple Silicon. Leverages unified memory for fast inference via MLX or Atomic Chat. Many models have official
LiquidAI/<model>-MLX-<bits>repositories; the others are community conversions on mlx-community. - ONNX β Best for production deployments and edge devices. Cross-platform with ONNX Runtime across CPUs, GPUs, and accelerators. Available as
LiquidAI/<model>-ONNXfor the LFM3 models and the main LFM2.5 text, vision, and audio models, and for some LFM2 models through onnx-community; see the matrix.
Quantization
Quantization reduces model size and speeds up inference with minimal quality loss. Available options by format:- GGUF β Text and vision repositories usually include
Q4_0,Q4_K_M,Q5_K_M,Q6_K,Q8_0,BF16, andF16. Some repositories ship a smaller set (for example, d1-omni-600M has onlyQ8_0,BF16, andF16), so check the file list.Q4_K_Moffers the best balance of size and quality. - MLX β Official LiquidAI repositories usually come in
4bit,5bit,6bit,8bit, andbf16. Some ship a subset (d1-omni-600M has8bitandfp16).8bitis recommended. - ONNX β Supports
FP32,FP16,Q4, andQ8(some repositories also includeQ4F16).Q4is recommended for most deployments.
Complete Model Matrix
Swipe horizontally to see all formats.
Looking for an older model? Deprecated models and their recommended replacements are listed on the Deprecations page.