inside
Post-training and inference for open models

Every domain has a shape. Your model should too.

Your own model. Trained on your service, served for less.

Talk to us →
01 · The problem

A generic model costs more, and still doesn’t know your service.

  • Consumer AIThe model your users love costs too much to serve at scale.
  • Agents & workflowsToo few tasks finish, and you pay for every one that doesn’t.
  • Regulated industriesYour data can’t leave, so the best AI is out of reach.
02 · Platform

We fit a model to your service and keep fitting it.

  1. 01LearnWe learn your domain from your data and what a good answer looks like to you.
  2. 02Post-trainWe fit an open model to your vocabulary, tone and rules.
  3. 03ServeWe run it at your cost and latency targets, or hand you the weights.
  4. 04ImproveReal usage becomes the next round of training.
03 · Consumer AI

Your users came for your product, not someone else’s model.

Character chatYour users prefer one model, and it’s expensive to serve.We tune your own model to feel the same, for a fraction of the cost.
Voice agentsYour callers hate waiting, and the API you use answers slowly.Your own model answers as soon as they finish.
A frontier API

Can you move my booking to nine? Sure, nine it is.

Your model

Can you move my booking to nine? Sure, nine it is.

04 · Agents & workflows

Not every task needs a frontier model.

Your agent sends every step to a frontier model, even the easy ones.Your own model does most steps for less. Trained on your work, it does some better.
What Coinbase found
“The frontier models are amazing generalists. But increasingly you’ll see specialist models out perform on more narrow tasks.”
Brian Armstrong, CEO of Coinbase
What we measuredQwen3.8 27B. $4.03 per task on the public leaderboard, $0.03–0.07 with our serving.Harvey’s legal agent benchmark (LAB)
05 · Regulated industries

Bring the model to your data.

Your data can’t leave, so the best AI is out of reach.We train your own model and run it inside your environment. Your data never leaves.
06 · Open-weight models

Any open model. Trained on your service, served for less.

Base model ● ready to tune
  1. DeepSeek V4.1 Flash
  2. GLM 5.3 Flash
  3. Kimi K3
  4. Qwen3.8 27B
  5. Gemma 4 31B
  6. Gemma 4 26B A4B
  7. DeepSeek V3.2
  8. gpt-oss 120B
  9. Qwen3 Omni 30B
  10. Muse Spark Open
DeepSeekMoEopen weights
Measured on customer traffic
62 msuntil a voice agent starts its replyCustomer’s target192 msWith Inside62 msGemma 4 31B · voice calls · p50
23 msuntil a speech model starts talking backStock serving213 msWith Inside23 msQwen3 Omni 30B A3B · speech to speech
−35%cost per character-chat requestOpenRouter price$1.00With Inside$0.65DeepSeek V3.2 · character chat
−65%GPUs for 2,200 live chats, at the same qualityTheir fleet before400 GPUsWith Inside138 GPUsGemma 4 26B A4B · role-play chat
All benchmarks and how we measure →
07 · Customers
“
At our scale, inference is the biggest line on the bill. Inside cut it by 70%, and our users didn’t notice a thing.
Case study: 2,200 live chats on 138 GPUs instead of ~400.Read the case study→
08 · Talk to us

Tell us about your service. We’ll show you its shape.

We measure on your traffic and your quality bar before you change anything.contact@fractalyze.io
What are you building?