Skip to content
TheChatArchiveThe Chat Archive
Perplexitysonar· PerplexityHugging Face LMArena Search Arena 24K

latest llm release

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with Geminigemini-2.0-flash-grounding’s answer.
  1. User

    latest llm release

  2. Perplexity Assistant

    The latest large language model (LLM) releases as of early 2025 include several notable advancements from major AI developers:

    • GPT-4.5 "Orion" by OpenAI: Released in February 2025, GPT-4.5 is OpenAI's largest model to date with an enormous 128,000 token context window. It outperforms GPT-4o in most benchmarks but is not designed as a reasoning model. It was first available to Pro users in late February 2025, with wider access for Teams, Enterprise, and Education users from March 2025. GPT-4.5 is expected to be the last major release before GPT-5[1].

    • DeepSeek R1 by DeepSeek: Released in January 2025, this reasoning-focused model has 671 billion parameters total (37 billion active) and a context window of 131,072 tokens. It excels in math and coding benchmarks, beating or matching OpenAI's o1 model. DeepSeek R1 is open-source and free to use via API and apps[1][5].

    • Llama 4 by Meta AI: Released April 5, 2025, Llama 4 includes variants like Llama 4 Scout and Llama 4 Maverick, which are multimodal models capable of processing text, video, images, and audio. Meta describes these as their most sophisticated and powerful models to date, with open-source availability. However, Meta reportedly postponed some aspects of Llama 4 due to concerns about reasoning and math performance compared to OpenAI's models[5][6].

    • Llama 3.3 by Meta AI: Released December 2024, Llama 3.3 is a 70 billion parameter multimodal model with a 128,000 token context window. It supports text and image inputs for complex analysis and is open-source, making it highly flexible for business applications[3].

    • Claude 3.7 Sonnet by Anthropic: Released early 2025, Claude 3.7 features an “extended thinking mode” for iterative reasoning and self-reflection, improving accuracy especially in coding and software development tasks. It is proprietary and known for strong summarization and conversational AI capabilities[3].

    • Gemini 2.5 by Google DeepMind: Released early 2025, Gemini 2.5 is a multimodal reasoning model with a massive 1 million token context window. It excels in complex problem-solving, coding, and multimodal content generation. Gemini 2.5 includes self-fact-checking features to improve reliability[3].

    • Mistral Small 3 by Mistral AI: Released January 2025, this 24 billion parameter model is optimized for low latency and high efficiency, suitable for real-time applications on limited hardware. It is open-source under Apache 2.0 license[3].

    Other notable models include OpenAI's o3 (successor to o1) released in December 2024, which focuses on analytical thinking and complex reasoning but is still in safety testing and not publicly available[8].

    Summary Table of Key Latest LLM Releases

    ModelDeveloperRelease DateParameters (B)Context Window (Tokens)Key FeaturesAvailability
    GPT-4.5 "Orion"OpenAIFeb 2025Unknown128,000Largest OpenAI model, broad knowledgePro, Teams, Enterprise
    DeepSeek R1DeepSeekJan 2025671 (37 active)131,072Reasoning, math & coding focusOpen-source, free
    Llama 4Meta AIApr 5, 2025~400UnknownMultimodal (text, video, images, audio)Open-source
    Llama 3.3Meta AIDec 202470128,000Multimodal, large open-source modelOpen-source
    Claude 3.7 SonnetAnthropicEarly 2025UnknownUnknownExtended thinking mode, coding focusProprietary
    Gemini 2.5Google DeepMindEarly 2025Unknown1,000,000Multimodal, self-fact-checkingProprietary
    Mistral Small 3Mistral AIJan 202524UnknownLow latency, efficient, open-sourceOpen-source
    OpenAI o3OpenAIDec 2024UnknownUnknownAnalytical thinking, complex reasoningTesting, not public

    These models reflect rapid innovation in LLM capabilities, including larger context windows, multimodal inputs, improved reasoning, and open-source availability for some. The field continues to evolve quickly with new releases expected throughout 2025[1][3][5][6][8].

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.