Skip to content
The Chat Archive
gemma-3-27b-it· Google LMArena Human Preference 140K

Tell me something about the best open LLM models in the planet.

  1. User

    Tell me something about the best open LLM models in the planet.

  2. Assistant

    Okay, let's dive into the world of top open Large Language Models (LLMs). "Best" is subjective and depends on what you need the model to do, but here's a breakdown of leading contenders as of late 2023/early 2024, categorized by their strengths, along with key details. I'll focus on models that are genuinely open – meaning weights are available, allowing for self-hosting and modification. I'll also give a sense of their licensing. Keep in mind this field is moving very fast!

    Important Considerations Before We Start:

    • Open vs. Open-Weight: There's a spectrum. Some models are "open-weight" (weights are available) but have restrictions on commercial use. Others are more truly open with permissive licenses.
    • Parameter Count: Generally, more parameters can mean more capability, but it's not the only factor. Architecture, training data, and fine-tuning are crucial.
    • Hardware Requirements: Larger models need significant GPU resources to run effectively.
    • Quantization: Techniques like quantization reduce the memory footprint of models, making them runnable on less powerful hardware, but can sometimes impact performance.
    • Fine-tuning: Most open LLMs benefit greatly from fine-tuning on a specific dataset for your use case.

    Here's a rundown of the top contenders, grouped by capability/focus:

    1. The Current Leader (Generally): Mistral 7B & its Derivatives

    • Model: Mistral 7B (and now Mistral 8x7B - Mixtral)
    • Developer: Mistral AI (French startup)
    • Parameter Count: 7 Billion (Mixtral is effectively 47B due to its Mixture of Experts architecture)
    • License: Apache 2.0 (Very permissive – good for commercial use)
    • Strengths: Exceptional performance for its size. Outperforms many larger models (like Llama 2 13B) on many benchmarks. Excellent reasoning, coding, and general knowledge. Mixtral is a huge leap forward, rivaling or exceeding GPT-3.5 level performance in many areas. Fast inference speed.
    • Weaknesses: Still susceptible to hallucinations (generating incorrect information). Can be sensitive to prompt engineering. Mixtral requires more VRAM than 7B.
    • Where to find it: Hugging Face Hub (https://huggingface.co/mistralai)
    • Derivatives: Many fine-tunes exist, like OpenHermes, Zephyr, and others, specializing in chat/instruction following.

    2. Llama 2 (Meta)

    • Model: Llama 2 (7B, 13B, 70B parameters)
    • Developer: Meta
    • License: Custom (Generally allows commercial use, but with restrictions – see Meta's terms. Larger companies may need to request a license.)
    • Strengths: Strong all-around performance. The 70B version is very capable and competitive with closed-source models like GPT-3.5. Good for a wide range of tasks. Large community support. Well-documented.
    • Weaknesses: The license is not as permissive as Apache 2.0. The 7B and 13B versions are less competitive than Mistral 7B. Can still hallucinate.
    • Where to find it: Hugging Face Hub (https://huggingface.co/meta-llama) and Meta's website.
    • Derivatives: A massive number of fine-tunes exist, including:
      • Vicuna: Early popular fine-tune for chat.
      • WizardLM: Focuses on complex instruction following.
      • Nous Hermes: Another strong chat-focused fine-tune.

    3. Falcon (Technology Innovation Institute)

    • Model: Falcon 7B, Falcon 40B, Falcon 180B
    • Developer: Technology Innovation Institute (Abu Dhabi)
    • License: Apache 2.0 (Permissive)
    • Strengths: Falcon 180B was, for a time, the leading open LLM. Falcon 40B is a strong performer. Trained on a massive dataset. Good for general-purpose tasks.
    • Weaknesses: Falcon 180B requires very significant hardware. Falcon 7B is less competitive than Mistral 7B. Can be prone to verbosity.
    • Where to find it: Hugging Face Hub (https://huggingface.co/tiiuae)

    4. Phi-2 (Microsoft)

    • Model: Phi-2
    • Developer: Microsoft
    • Parameter Count: 2.7 Billion
    • License: Microsoft Research License (Allows research and non-commercial use. Commercial use requires a license.)
    • Strengths: Remarkable performance for its tiny size. Outperforms models many times its size on some benchmarks. Excellent reasoning and language understanding. Can run on very modest hardware.
    • Weaknesses: Smaller size means it has limitations in knowledge and complexity. License is restrictive for commercial applications.
    • Where to find it: Hugging Face Hub (https://huggingface.co/microsoft/phi-2)

    5. OpenLLaMA (Berkeley AI Research)

    • Model: OpenLLaMA (3B, 7B, 13B)
    • Developer: Berkeley AI Research
    • License: Apache 2.0
    • Strengths: Aims to be a truly open reproduction of Llama. Good for research and experimentation.
    • Weaknesses: Generally lags behind Llama 2 and Mistral in performance.

    Here's a quick table summarizing the key points:

    ModelParametersLicenseStrengthsWeaknesses
    Mistral 7B7BApache 2.0Excellent performance for size, fastHallucinations, prompt sensitivity
    Mixtral 8x7B~47BApache 2.0Near GPT-3.5 level, very capableHigher VRAM requirement
    Llama 27B, 13B, 70BCustomStrong all-around, large communityLicense restrictions, 7/13B less competitive
    Falcon7B, 40B, 180BApache 2.0Massive dataset, good general purpose180B needs huge hardware, verbosity
    Phi-22.7BMicrosoft ResearchTiny size, excellent reasoningLimited knowledge, restrictive license
    OpenLLaMA3B, 7B, 13BApache 2.0Truly open reproduction of LlamaLower performance than Llama 2/Mistral

    Resources for Staying Up-to-Date:

    To help me narrow down the best recommendation for you, could you tell me:

    • What do you want to do with the LLM? (e.g., Chatbot, code generation, text summarization, creative writing, research)
    • What hardware do you have available? (e.g., GPU with how much VRAM, CPU only)
    • What is your licensing requirement? (e.g., Commercial use allowed, research only)

    I hope this comprehensive overview is helpful! Let me know if you have any other questions.

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