Show HN: I estimated the carbon impact of different LLMs
15 days ago
- #CO2-emissions
- #AI-models
- #environmental-impact
- Comparison of environmental impact of popular AI models based on CO₂e emissions, energy consumption, and cost per 1M tokens.
- Rankings include models from providers like moonshotai, meta-llama, z-ai, openai, mistralai, anthropic, google, deepseek, qwen, baidu, and x-ai.
- Top-ranked model: moonshotai/kimi-vl-a3b-thinking with lowest CO₂e emissions (~0.00006232 g per 1M tokens) and energy consumption (~119.9mWh per 1M tokens).
- Highest CO₂e emissions: mistralai/mistral-medium-3.1 (~4.17 g per 1M tokens) and highest energy consumption (~9259.7Wh per 1M tokens).
- Cost per 1M tokens varies widely, from $0.018 (meta-llama/llama-3.1-8b-instruct) to $50.00 (openai/o3-pro).
- Environmental impact estimates are based on calibrated models and may vary based on hardware, data center efficiency, and energy sources.
- Encouragement for providers to share accurate data to improve estimates.