AI Models Database

Dense hybrid Qwen 3.5 model with strong reasoning and multimodal capabilities. Best quality-to-size ratio in the family for single-GPU deployment.

Sub-second image generation on consumer GPUs. Unified text-to-image and image editing in one checkpoint. Apache 2.0 licensed for commercial use. Best open fast image model.

Z.ai flagship open-source MoE model with 358B parameters, 200K context window, and 128K output. Excels at multilingual tasks, coding, and reasoning.

Allen AI OLMo 3 32B - large scale fully open model. Complete transparency with training data, code, and intermediate checkpoints.

DeepSeek V3.2 with 685B total parameters (37B active). Features Multi-Token Prediction and Sparse Attention for efficiency.

Arcee Trinity Mini with 26B total parameters but only 3B active (128 experts). Extremely efficient MoE architecture.

IBM Granite 4.0 32B flagship model. Hybrid Mamba-Transformer for efficient long-context processing.

Chain-of-thought reasoning variant of Phi-4 Mini (3.8B). SFT on OpenAI o3-mini demonstrations + RL. 128K context makes it ideal for RAG. Best reasoning at sub-4B scale.

Efficient MoE multimodal model with industry-leading 10M token context window. 109B total / 17B active parameters across 16 experts. Native text + image understanding. Trained on 40T tokens in 12 languages.

Largest Gemma 3 model with multimodal vision capabilities. 128K context, 140+ languages. QAT variants available for 3x smaller footprint.

Multi-speaker dialogue TTS with non-verbal sounds (laughs, sighs, coughs). Voice cloning via audio prompt conditioning. Best model for scripted dialogue and podcast generation.

First open-source MoE text embedding model. 475M total / 305M active parameters. Matryoshka flexible dimensions (256-768). State-of-the-art multilingual embedding at release.

Mistral Small 3 with 24B parameters. Supports vision, function calling, and 32K context. Apache 2.0 licensed.

Ultra-lightweight TTS model. Under $1 per million characters. 54 pre-built voices across 8 languages. Apache 2.0 for commercial deployment. 8.9M+ monthly HuggingFace downloads.

Efficient local coding model. Optimized for code generation, completion, and development assistance.

Efficient 8B model with vision support. Good balance for local deployment with moderate resources.

4.5x faster than Whisper Large V3 with minimal quality loss. Decoder reduced from 32 to 4 layers. Most-downloaded Whisper variant (4.6M+ monthly). Best speed/accuracy balance for local STT.

01.AI coding model with 128K context

Fine-tuned Llama 3 without alignment restrictions

Mistral dedicated coding model, supports fill-in-the-middle

High-quality embeddings for semantic search

Cohere enterprise model optimized for RAG workflows

Tiny but capable vision model

BigCode latest coding model with improved performance

768-dimensional text embedding model optimized for semantic similarity search. ~1-2s per embedding on CPU. Used for ProHive RAG system.

BAAI embedding model for RAG applications

Larger LLaVA with improved visual reasoning

Defog SQL generation model, beats GPT-4 on SQL

Mixture of Experts model with 8 experts, uses 12B active params

Microsoft Orca with improved reasoning capabilities

Intel optimized chat model

Lightweight embedding model for quick inference