Methodology
Documented assumptions you can adjust
2,500 tok/GPU-s · 150W GPU · 55% KV share · 0.417 kg CO₂/kWh. Illustrative — not live metering.
- Tokens saved = original − kept per compression
- Full model in Environment guide
The new standard of context compression.
Every oversized prompt burns GPU cycles, grid power, and cooling water on tokens that never needed to run. SuperCompress cuts ~65% of context before inference — keeping what matters, lowering cost, and giving resources back.
~65% fewer tokens · query-aware · CPU-only policy
Every larger model turn pulls more from the grid — more power, more water for cooling, more strain on resources already running thin. We send millions of tokens through GPUs that never needed to be processed — and watch the cost add up in silence. We cannot scale intelligence by burning through what we have left. SuperCompress cuts wasted tokens before inference — keeping what matters, using less compute, and giving resources back to the planet.
Measure the compute you can skip
Paste real context and your question. SuperCompress drops everything it safely can — keeping only what matters for the answer.
Quality
Savings without dropping answers
Truncation can save tokens but lose critical lines. SuperCompress targets both KV reduction and oracle recall.
- 100% oracle recall vs ~25% baselines
- Failure case demo at budget 0.1
Deployment
CPU eviction before GPU inference
~5K-parameter network runs on CPU. Sub-ms latency vs much larger GPU prefill savings on long context.
- pip install · MIT license
- Live hosted API + dashboard
Try a preset or bring your own context
Enter in question field · ⌘/Ctrl+Enter in context
Runs locally with the trained SuperCompress policy. Your question defines what’s important — the engine evicts low-score lines until only answer-critical context remains.
Token & environmental impact
- Input tokens —
- Tokens after SuperCompress —
- Tokens removed —
- Tokens saved —
- Context size saved —
- Answer quality retained —
- Electricity (before) — est. prefill
- Electricity (after) — est. prefill
- Power saved —
- Water saved — datacenter cooling (est.)
- CO₂ saved —
Cut waste. Keep the answer.
On long contexts, adaptive mode typically removes 85–95% of tokens while keeping query-critical lines. At a fixed 35% budget, SuperCompress hits 100% oracle recall on benchmark seeds — FIFO and truncation retain ~25%.
Quick start
curl -X POST https://trysupercompress.vercel.app/api/v1/compress \
-H "X-API-Key: sc_live_…" \
-d '{"context":"…","query":"…"}'
Production API — metered in your dashboard. Requires a sc_live_… key.
Python: pip install git+https://github.com/arjunkshah/supercompress.git
· API reference