Which LLM is better?
- User
Which LLM is better?
- Assistant
“Better” really depends on what you need the model to do. Here are some quick guidelines, followed by a comparison table and a few questions to help you narrow things down.
-
Define the goal
• Casual chat or creative writing?
• Enterprise-grade reasoning or research?
• Code generation?
• On-device inference with tight latency or privacy constraints? -
Pick the metrics that matter to you
• Raw accuracy / reasoning ability
• Speed & cost
• Model size & deployability (cloud vs. local)
• Licensing / commercial terms
• Tooling & ecosystem support -
Short headline answers
• Best raw reasoning (today, June 2024): GPT-4o (OpenAI) and Claude 3 (Opus) are generally top of public leaderboards.
• Best cost-to-performance open-weights: Llama-3 70B, Mixtral 8×22B, or Mistral-7B v0.2 depending on the hardware you have.
• Easiest to fine-tune / run locally: Llama-3 and Mistral families (Apache 2 license).
• Strongest for code: GPT-4o, Claude 3 (Opus/Sonnet) and specialized variants like DeepSeek-Coder or Code Llama.
• Multimodal (image + text): GPT-4o and Gemini 1.5 Pro lead in publicly available models.
Popular LLMs at a glance (June 2024)
Model Provider / License Params Strengths Weak spots Typical cost & access GPT-4o OpenAI (closed) ~Trillions (Mixture-of-Experts) SOTA reasoning, multilingual, vision, audio; rich ecosystem Closed weights, usage cost, no local deployment $0.005-0.01 / 1K tokens (chat); API or ChatGPT Claude 3 (Opus) Anthropic (closed) undisclosed MoE Long-context (200K+ tokens), strong reasoning, safe outputs Closed weights, priced higher than GPT-4o $15 / 1M input tokens Gemini 1.5 Pro Google (closed) undisclosed MoE 1M-token context, good multimodal, code Limited public access, slower roll-out Vertex AI, AI Studio Llama-3 70B Meta (Apache-2) 70 B Open weights, robust chat, easy fine-tune Needs high-end GPU for real-time chat; behind GPT-4o on complex tasks Free weights; infra cost only Mixtral 8×22B Mistral (Apache-2) 8 experts, 22 B each Great throughput/price, near-GPT-3.5 quality, open weights Lags GPT-4-class reasoning Hugging Face / Mistral API Mistral-7B-v0.2 Mistral (Apache-2) 7 B Runs on laptop/phone, solid quality for size Limited reasoning depth Free weights DeepSeek-Coder 33B DeepSeek (Apache-2) 33 B Code generation & completion General chat weaker Free weights Command-R Cohere (closed) undisclosed Retrieval-augmented design, enterprise docs Closed weights Cohere API
Questions to pinpoint “better” for you
- Deployment: Do you need the model on-prem / offline, or is cloud fine?
- Budget: Are you optimizing for cost per 1 K tokens, or is quality paramount?
- Domain: Will it be used mostly for code, legal text, marketing copy, or general Q&A?
- Context length: Do you need 16K, 128K, or 1M-token windows?
- Compliance: Any data-sovereignty or licensing constraints?
Feel free to share a bit more about your use-case, and I can give a sharper recommendation.
-
Other models, same or similar prompt
Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.