surogate/surogate3.7-flash

Surogate 3.7 Flash.

Our own model for agent work. It reads text and images, calls your tools and holds a million tokens, at a price that lets agents run all day.

Book a demoTry it on llm.buzz

Listed on llm.buzz since Oct 20, 2026.

POST /chat/completions · surogate/surogate3.7-flash

Ready

Check this delivery note against the order. Anything off?

delivery-note.jpgPO-88213.pdf

First token
·
Speed
·
Tokens
0 in · 0 out
This request
$0.00000

One request, played at the speed and prices llm.buzz lists for it. A demo order.

tokens of context, 262,144 native and stretched four times with YaRN
1M
tokens of context, 262,144 native and stretched four times with YaRN
tokens in one answer
32K
tokens in one answer
to the first token, then 112 tokens a second, on llm.buzz
0.04 s
to the first token, then 112 tokens a second, on llm.buzz
per million input tokens, $1.00 per million out
$0.15
per million input tokens, $1.00 per million out

  • Text and images in

    Photos of delivery notes, scanned forms, screenshots and charts, read alongside the text.

  • Tools

    It calls your functions mid-answer and works with the result, for as many turns as the job takes.

  • Structured outputs

    Answers in the JSON schema you give it, so your code reads the result without parsing prose.

  • Prompt caching

    A long system prompt or document is billed at $0.05 per million tokens after the first read.

Head to head with frontier models.

Our model runs on your own servers and scores level with or ahead of each of these on real agent tasks.

  1. Surogate 3.7 FlashInvergent, open weights
    54.6
  2. Kimi K3Moonshot AI, open weights
    54.5
  3. GLM 5.2Zhipu AI, open weights
    54.2
  4. Claude Opus 4.6Anthropic, cloud only
    51.6
  5. GPT-5.4OpenAI, cloud only
    50.3
  6. DeepSeek V4 ProDeepSeek, open weights
    43.7
  7. Gemini 3.1 ProGoogle DeepMind, cloud only
    40.8
Score in percent on 60 agent tasks.WildClawBench, read Oct 29, 2026

A million tokens, and it still finds the line.

A whole tender with its annexes, a year of a customer’s emails, or an agent’s long working session, in one prompt. Planted a single fact at a full million tokens, it found it every time.

NativeStretched with YaRN, factor 4
0262,1441,048,576 tokens

Faster on our own engine.

Served by the Surogate engine in NVFP4 on one NVIDIA DGX Spark, against vLLM on the same machine.

Tokens a second, and agent turns a minute

  • Decode, 1 user1.16× vs. vLLM
    Surogate70.4 vLLM60.8
  • Decode, 32 users1.17× vs. vLLM
    Surogate473.6 vLLM406.4
  • 8 agents at 100k to 200k tokens each, turns a minute1.85× vs. vLLM
    Surogate15.9 vLLM8.6
to the first token of a cold 108k-token prompt, against 32.4 s on vLLM
22.2 s
to the first token of a cold 108k-token prompt, against 32.4 s on vLLM
to the first token with 8 agents on warm caches, against 7.6 s on vLLM
1.9 s
to the first token with 8 agents on warm caches, against 7.6 s on vLLM
needle tests passed at a full million tokens
3 of 3
needle tests passed at a full million tokens

From the engine’s README.

Served from the EU.

Two providers serve it on llm.buzz today. Neither trains on your prompts, and neither keeps them.

ProviderInputOutputFirst tokenSpeed
Surogate SparkEU · Does not train on prompts and keeps nothing.$0.15$1.000.04 s112 tok/s
Simpleminds EUEU, FP4 · Does not train on prompts and keeps nothing.$0.10$0.750.07 s134 tok/s

Per million tokens, as llm.buzz lists them.

Call it like any other model.

The OpenAI and Anthropic SDKs work as they are. Set the base URL, then call it by name.

AI Gateway 
import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.llm.buzz/api/v1",
    api_key=os.environ["LLM_BUZZ_KEY"],
)

reply = client.chat.completions.create(
    model="surogate/surogate3.7-flash",
    messages=[{"role": "user", "content": "Hello"}],
)

Rent it by the token, or own it.

  • On llm.buzz

    Pay per token, from the EU. Pick the provider, or let the router pick by price or speed.

    Open on llm.buzz

  • On your servers

    With Surogate Enterprise, on your own GPUs, served by our engine, with nothing sent to us.

    Talk to us

  • Trained on your work

    We fine-tune it on your documents and decisions, evaluate it on your own questions, and you keep the version you pin.

    Custom models

Questions

What is Surogate 3.7 Flash?

Our own model for agent work: it reads text and images, calls tools, answers in your JSON schema and holds a million tokens of context. We train it in our lab and serve it from the EU.

How do I try it?

On llm.buzz, with any OpenAI-compatible client: set the base URL and call surogate/surogate3.7-flash. Chat Completions, Responses and Messages endpoints all work.

Where does my data go?

Both providers that serve it on llm.buzz run in the EU, keep prompts for zero days and do not train on them. On your own servers, nothing leaves your network.

How was the 54.6 measured?

With WildClawBench’s own harness on its 60 agent tasks, run by us. The other scores are the public leaderboard’s, read on Oct 29, 2026.

Can it run on our own hardware?

Yes, with Surogate Enterprise. It is served by our engine, which is also what makes it fast on a single DGX Spark.

Can you train it on our data?

Yes. We fine-tune it on your work, score it on questions your data never asked, and serve the version you choose.

Put it to work on yours.

We will show it on your documents, in your cloud or on your hardware.

Book a demo