AI that thinks about
the planet

GreenInfer routes every prompt to the most energy-efficient model capable of answering it accurately, cutting AI compute energy by up to 97% vs always using the largest model.

Try GreenInfer How it works
97%
Max Energy Saved
73%
Avg Reduction
98.9%
Classifier Accuracy
183g
CO₂ Avoided
Pipeline

How GreenInfer works

Every prompt passes through a 7-layer orchestration pipeline that scores complexity, optimizes tokens, and routes to the right model tier before a single GPU cycle runs.

01
📝
Prompt In
User sends query
→
02
🔬
Complexity Score
0 to 100 via DistilBERT
→
03
⚡
Prompt Optimize
T5 removes filler tokens
→
04
🌎
Carbon Route
ERCOT grid check
→
05
🎯
Model Route
Small / Medium / Large
→
06
✓
Answer
With energy metrics
Small TierEco
0.9 mWh
Llama 3.2 1B  ·  Simple queries  ·  55% of traffic
Medium TierBalanced
3.8 mWh
Llama 3.1 8B  ·  Reasoning  ·  30% of traffic
Large TierHeavy
48.0 mWh
Llama 3.3 70B  ·  Complex / Code  ·  15% of traffic
Routing examples
"What is the capital of France?"→
Small0.9 mWh · -98%
"Explain transformer attention mechanisms"→
Medium3.8 mWh · -92%
"Build a REST API with JWT authentication in Python"→
Large48 mWh · baseline
Features

Built for sustainable inference

Every layer reduces unnecessary compute while preserving answer quality.

🔬
Complexity Scoring
DistilBERT classifier trained on 600 labeled examples rates every prompt 0 to 100. 98.9% validation accuracy across 4 tiers.
⚡
T5 Prompt Optimizer
Silently rewrites prompts to remove filler words before inference, cutting input tokens by an average of 35% with no quality loss.
🌎
Carbon-Aware Routing
Uses hourly ERCOT grid intensity estimates to defer expensive queries when the grid is running dirty, reducing CO₂ further.
🤺
Cascade Engine
Starts with the smallest model and only escalates to medium or large if confidence is below threshold, inspired by FrugalGPT.
💡
Smart Preview Mode
For complex queries, shows a summary and outline first. User confirms before the full expensive response runs, preventing wasted reruns.
📊
Per-Response Metrics
Every answer shows energy used in mWh, CO₂ emitted, tokens saved, and a Green Efficiency Score 0 to 100 with an improvement tip.
Benchmark Results

Real numbers from real experiments

All figures from running the benchmark suite on 20 prompts across complexity tiers. Source: experiments/benchmark.py

97%
Energy saved (small-tier queries)
vs always using Llama 3.3 70B
73%
Average energy reduction
mixed workload, all query types
98.9%
Classifier accuracy
DistilBERT, 600 training examples
183g
CO₂ avoided
20-prompt benchmark test set
55%
Routing accuracy
matching human-labeled tier
35%
Avg token reduction
via T5 prompt optimizer
Python Framework

Drop-in green inference

Import the package and get energy-aware responses in three lines.

Python — greeninfer package
from greeninfer import GreenInfer # Auto-routes to the right model tier gi = GreenInfer() result = gi.chat("Explain photosynthesis") print(result.response) # "Photosynthesis is the process by which..." print(result.energy_mwh) # 0.9 (small model, 98% less than large) print(result.model_tier) # "small" print(result.co2_grams) # 0.000178 # Switch modes result = gi.chat("Write a sorting algorithm", mode="eco") result = gi.chat("Analyze this legal contract...", mode="performance")
Full Framework Docs

Ready to chat green?

Every message shows exactly how much energy it used and how much was saved by not using the largest model.

Open Chat View Impact Data