Pareton

Blog

Notes from the frontier.

Talks, results, and what we learn making inference faster and cheaper on real workloads.

Xavier Lyu on stage at Exploit Summit, in front of a slide titled Compounding Inference Optimization

Field notes

The pit crew, not the engine

Notes from Xavier Lyu's talk at Exploit Summit in Montréal: why inference gains compound, and why they only count on the workload you already run.

Read · 5 min

Every token, priced live. Pareton is now a verified provider on the Liquid Inference network.

Network

Pareton joins Liquid Inference as a verified provider

Inference is turning into a market with live prices. On a market like that, cost per token is the whole game.

Partnership: Engy is getting faster. Pareton now tunes how Engy runs its AI models.

Partnership

Engy is getting faster

Three rounds into our Kimi K3 campaign, the leading patch serves Engy's traffic about 25% faster end-to-end, on the same eight B300 GPUs.

Chart: Qwen3.8-27B on SGLang and four RTX 5090s. Output throughput versus baseline: +54% at 4 concurrent requests, +50% at 8, +40% at 16, +23% at 32.

Results

Same model. Same GPUs. Up to 54% more tokens per second.

We benchmarked the top miner patch for Qwen3.8-27B against the pinned baseline with NVIDIA's AIPerf. Output throughput rose 23 to 54% across every load we tested.

Diagram comparing TP=4 and TP=8: model weights and KV cache are distributed across GPUs, temporary memory per GPU stays the same.

Engineering

More GPUs don't always mean more usable memory

A workload hit an out-of-memory error. Doubling the GPUs would not have fixed it. Here is why, and what did.

Pushing the Pareto frontier of inference.

If you already serve models under a real SLA, send the profile. We will tell you whether the search is worth running.

Talk to usOpen the dashboard