LLM Models

Suno's AI music generation model (May 2025), producing full songs with vocals and instrumentation from a text prompt. v4.5 extended song length to 8 minutes, added 1,200+ genre tags, and improved vocal expressiveness and prompt adherence for coherent, radio-quality tracks.

Meta's AI music generation model available in 300M, 1.5B, and 3.3B parameter sizes. Single-stage auto-regressive Transformer trained over 32kHz EnCodec tokenizer with 4 codebooks sampled at 50Hz. Trained on 20K hours of licensed music from internal dataset (10K high-quality tracks) plus ShutterStock and Pond5 collections (390K instrument-only tracks). Eliminates cascading model requirement through efficient token interleaving - only 50 auto-regressive steps per second of audio. Supports both text-to-music and melody-guided generation. Evaluated on MusicCaps benchmark showing superiority vs baselines. Released April-May 2023. Part of AudioCraft toolkit alongside AudioGen and EnCodec.

Jina AI's open text embedding model (September 2024), a 570M-parameter model with task-specific LoRA adapters (retrieval, clustering, classification) and long-context support. It offers strong price-performance and multilingual retrieval, self-hostable or via the Jina API.

Nomic's open text embedding model (February 2024), fully open with public weights, training code, and data. It offers an 8,192-token context - among the longest for open models - with Matryoshka dimensions, competitive quality, and reproducible, auditable training.

BAAI's open-source multilingual embedding model, notable for Multi-Functionality (dense, sparse, and multi-vector retrieval), Multi-Linguality (100+ languages), and Multi-Granularity (inputs up to 8,192 tokens). It is the strongest open option for multilingual search, matching commercial quality when self-hosted.

Google's embedding model that maps text (and, in multimodal variants, images, video, audio, and PDFs) into a shared vector space, leading retrieval benchmarks (~67.7 MTEB retrieval). It supports Matryoshka dimensions up to 3,072 and integrates with the Gemini API and Vertex AI.

Voyage AI's retrieval-optimized commercial embedding model, among the best for MTEB retrieval quality at production scale. It supports long context and Matryoshka (shortenable) dimensions, and is a common choice for high-accuracy RAG pipelines.

Cohere's multimodal embedding model, a top MTEB performer that embeds text and images into a shared space with support for 100+ languages and long documents. Built for enterprise search and RAG, it pairs with Cohere Rerank for high-precision retrieval pipelines.

OpenAI's largest text embedding model (January 2024), producing up to 3,072-dimensional vectors for retrieval, RAG, clustering, and classification. It supports shortening dimensions via the API for storage/latency trade-offs and remains a strong general-purpose retrieval baseline.

OpenAI's state-of-the-art automatic speech recognition (ASR) model with 1.55B parameters supporting 99+ languages. Trained on 1M hours of weakly labeled audio plus 4M hours of pseudo-labeled audio from Whisper large-v2. Uses 128 Mel frequency bins (vs 80 in previous versions) and trained for 2.0 epochs. Achieves 7.4% WER average and 97.9% word accuracy on LibriSpeech. Shows 10-20% error reduction vs large-v2 and 72% WER reduction vs prior MLPerf ASR model (RNN-T). Performs automatic language identification, generates phrase-level timestamps, and handles punctuation/capitalization. Strongest on high-resource languages (English, Spanish, French, German). Open source enables self-hosting and commercial use.

Technology Innovation Institute's massive 180B parameter open-access model trained on 3.5T tokens. Causal decoder-only architecture with 80 layers, hidden dimension 14,848, vocabulary size 65,024. Trained on up to 4,096 A100 GPUs using Amazon SageMaker for ~7M GPU hours. Dataset consists of 85% RefinedWeb plus curated conversations, technical papers, and code (~3%). Achieved 68.74 on Hugging Face Open LLM Leaderboard - highest among open models at release. Surpassed Meta's LLaMA 2 and ranks near GPT-4 and PaLM 2. 2.5x larger than Llama 2 with 4x more compute. Released under Falcon 180B TII License (Apache 2.0 based) for research and commercial use.

BigCode's open-source code generation model with 15B parameters trained on 4T+ tokens from The Stack v2 dataset covering 600+ programming languages. Features 16,384-token context window enabling handling of longer codebases and elaborate instructions. Trained on 1T tokens of permissively licensed GitHub data including code, Git commits, issues, and Jupyter Notebooks. Achieves 33.6%-44.2% on HumanEval and excels on MultiPL-E across 16 of 18 programming languages. Best-performing large model on DS-1000 benchmark. Outperforms CodeLlama-34B on math and code reasoning while matching low-resource language performance. Released under permissive BigCode OpenRAIL-M license. Collaboration between ServiceNow, Hugging Face, and NVIDIA.