Models46 articles
Llama 4 Scout
Llama 4 Scout is a high-efficiency multimodal large language model released by Meta AI in April 2025, utilizing a mixture-of-experts (MoE) architecture with 109 billion total parameters. It is distinguished by its massive 10-million-token context window and its ability to natively process and integrate text and image inputs.
V3.1
V3.1 is an open-weights large language model developed by DeepSeek that integrates general-purpose conversational capabilities with advanced reasoning features. It utilizes a 671-billion parameter Mixture-of-Experts architecture and features a hybrid thinking mode allowing for chain-of-thought processing.
V3.2 Exp
V3.2 Exp is an experimental large language model developed by DeepSeek that introduced the DeepSeek Sparse Attention (DSA) mechanism to optimize long-context processing. Released in September 2025, it utilizes a Mixture-of-Experts architecture and serves as a hybrid model for both general-purpose instruction and complex reasoning.
Phi-4 Multimodal
Phi-4-reasoning-vision-15B is a 15 billion parameter multimodal model developed by Microsoft Research that integrates visual and audio perception with structured reasoning. Released in March 2026, it utilizes a mid-fusion architecture to perform complex tasks like UI grounding and scientific problem-solving on modest hardware.
Qwen 3 14B
Qwen 3 14B is a dense 14.8 billion parameter large language model developed by Alibaba Cloud, featuring a hybrid reasoning engine that allows toggling between thinking and non-thinking modes. Released in April 2025 under an Apache 2.0 license, it achieves performance parity with much larger previous-generation models in STEM, coding, and logical reasoning tasks.
Qwen 3 30B A3B
Qwen 3 30B A3B is a 30-billion parameter large language model developed by Alibaba Cloud, utilizing a sparse Mixture-of-Experts (MoE) architecture that activates 3 billion parameters per token for high efficiency. It is designed to provide mid-sized model reasoning capabilities with the speed and cost profile of a much smaller system, supporting a context window of up to 262,144 tokens.
R1-0528 Turbo
R1-0528 Turbo is a high-efficiency large language model developed by DeepSeek AI, optimized for throughput and complex reasoning using a Mixture-of-Experts (MoE) architecture. It is designed to provide advanced logic and coding capabilities with significantly reduced computational overhead and API costs.
Sonar
Sonar is a family of large language models developed by Perplexity AI, specifically optimized for Retrieval-Augmented Generation (RAG) and real-time search synthesis. Built on Meta's Llama architecture, the models prioritize factual groundedness and source attribution to power Perplexity's 'answer engine' platform.
Sonar Pro
Sonar Pro is a search-centric large language model developed by Perplexity AI, built on the Llama 3.3 70B architecture to provide high-speed, fact-grounded responses with real-time internet connectivity.
V3-0324
DeepSeek V3-0324 is an open-weights Mixture-of-Experts (MoE) language model with 671 billion total parameters, optimized for high-level reasoning, coding, and technical tasks. Released in March 2025, it features innovations like Multi-head Latent Attention and Multi-Token Prediction to balance high performance with inference efficiency.