Centauryentaury
Research · FauzenResearch

A small model that reasons like a big one.

The Fauzen family bets that a disciplined reasoning process beats raw size — powerful, fast, and cheap. This is our research frontier, not a product you can buy yet.

The bet

Most labs chase capability by adding parameters. We chase it by making a smaller model think in disciplined steps — plan, reason, verify — before it answers.

01

Reasoning, not memorizing

A structured chain — decompose, solve, check — lets a compact model match far larger ones on problems that reward thinking over recall.

02

Cheap to run

Fewer parameters means less GPU per answer. The reasoning budget is spent where it changes the outcome, not burned on every token.

03

Fast where it counts

Short problems take a short path; only the hard ones open the full reasoning loop. Latency scales with difficulty, not with model size.

How it stacks up

Illustrative only — placeholder numbers until public evals land. We won't publish a benchmark we can't reproduce.

Reasoning eval · higher is better

Mock data · pending public benchmarks
  • Fauzen R188
  • Frontier-A91
  • Frontier-B85
  • Open-70B74
  • Open-8B52

cheaper per task than a frontier model

faster median response on hard problems

0/10

the parameter count of the frontier tier

The family

The same reasoning loop, distilled to fit the job — from deep research to real-time.

Fauzen R

Reasoning

The full loop. Deep chains, self-verification, tool use — for the problems where being right matters more than being instant.

  • Deep chains
  • Self-verify
  • Tool use

Fauzen Mini

Balanced

The everyday model. Enough reasoning to stay sharp, distilled for cost — the one you'd wire into a product.

  • Distilled
  • Low cost
  • General

Fauzen Flash

Fast

Reasoning on a budget. Short path for easy asks, full loop only when the problem earns it. Built for real-time.

  • Adaptive path
  • Real-time
  • Batch-ready

Straight talk

This is research, not a release.

Fauzen is where we investigate — training runs, ablations, evals we can actually reproduce. We won't sell you a model that isn't ready, and we won't dress a placeholder chart up as a benchmark. When there's something real to ship, you'll see the numbers first.

Follow the work

Want to track the research, or put reasoning-first models to work when they ship? Tell us what you're building.