A model you own, on your infrastructure

Surogate 3.7 Flash performs level with frontier models on enterprise tasks. It runs in your cloud account, on your own servers or fully offline.

Banking · On your servers

A record of how each case was decided

Agents work through credit files, KYC packs and complaints on your own servers. For each case they log the documents they read, what they did and who signed off, so a decision can be explained to a regulator years later.

Explore

surogate/surogate3.7-flash

Surogate 3.7 Flash, our model for agentic work

Trained in our lab. It holds a million tokens, sees your documents, calls your tools and scores level with frontier models on enterprise tasks, on our EU cloud or on your own hardware.

A million tokens of contextA full tender in one prompt

262,144 tokens native, stretched four times with YaRN. In our test, a single fact placed at one million tokens was found 3 times out of 3.

1Mtokens of context

The model

Text and images inScans, photos and charts included

Scans, photos of delivery notes, screenshots and charts, read alongside the text, with answers up to 32K tokens long.

32Ktokens in one answer

What it can do

It calls your toolsAnd answers in your schema

It calls your functions mid-answer, works with what comes back, and answers in the JSON schema you give it. Long prompts are cached.

3APIs: Chat Completions, Responses and Messages

Call it

Level with frontier modelsOn enterprise tasks

WildClawBench: 60 agent tasks, scored with the benchmark's own harness, in our run.

54.6percent on WildClawBench

All results
  1. Surogate 3.7 Flash54.6
  2. Kimi K354.5
  3. GLM 5.254.2
  4. Claude Opus 4.651.6
  5. GPT-5.450.3
  6. DeepSeek V4 Pro43.7
  7. Gemini 3.1 Pro40.8

Fast on our own engineAffordable for continuous workloads

On one DGX Spark it gives 8 long-running agents almost twice the turns vLLM does. From the EU at $0.15 per million tokens in, or on your own GPUs.

1.85×agent turns a minute, against vLLM

Try it on llm.buzz
  • Decode, 1 user1.16×

  • Decode, 32 users1.17×

  • 8 agents at 100k to 200k tokens each, turns a minute1.85×

  • Surogate enginevLLM, same machine

Already building AI solutions for your clients?

Through the partner programme, you deliver your work on Surogate, Keel and custom models. You receive a share of what each client pays for as long as the client stays, and we can train your team.

  • No fee to join
  • Your workspace has no plan fee
  • 60 days free for every new client
  • 30% every month, 40% at Diamond

Your practice

Silver partner, 0 of 4 clients live

30%

your share

Claims intake

One agent, built and tested once in your own workspace.

  • An insurer

    On its own servers

    Next
  • A hospital group

    Air-gapped, on site

    Next
  • A city hall

    On our EU cloud

    Next
  • A freight forwarder

    In its own cloud account

    Next

The full stack, not only a model

Surogate, Surogate Desktop, Keel, the AI Gateway and models trained on a client’s own data, on our cloud, on the client’s own Kubernetes, or air-gapped.

A share for as long as the client stays

30% of each client’s plan or licence, every month they stay, with no cap at twelve months. Diamond partners earn 40%.

Tiers follow live clients

There is no application. Your tier is set by the number of clients you have live. Gold partners are listed in the directory and introduced to companies looking for a partner.

The first 10 partners get Diamond terms for their first 12 months, whatever their tier.

Walk through one process with us.