Applied research, shipped.

Ogbon is an applied research lab working on fast, low‑cost AI. We run the experiments, publish what we find, and turn what works into products you can use.

Try the platform Read our research

01 / The lab

One question, every result public.

How do you make AI models fast and cheap enough to answer in real time? We test ideas, measure them, and build on the ones that work.

  • Measure first

    Every number comes with its raw data.

  • Publish every result

    Including the experiments that failed.

  • Build on what works

    What works goes into our products.

02 / What we’re building

A platform and a model.

Developer preview

Ogbon Platform

Inference serving for open models. Call managed models through one API, or deploy your own from Hugging Face.

  • OpenAI‑compatible API
  • Managed and custom models
  • Pay per token from a prepaid wallet
Get started How it works

In development

ogbon‑1

Our first model. We are building it together with the serving platform.

  • Built by the Ogbon lab
  • Served on Ogbon Platform
  • Details at release
Get updates About ogbon‑1

03 / Ogbon Platform

Switch in one line.

Use the OpenAI SDK you already have. Change the base URL and pick a model.

app.py

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.ogbon.com/v1",
    api_key=os.environ["OGBON_API_KEY"],
)

reply = client.chat.completions.create(
    # A managed model, or one you deployed
    model="llama-3.1-8b",
    messages=history,
)

  • Managed models

    Popular open models, hosted by us.

  • Your own models

    Deploy any model from Hugging Face.

  • Prepaid wallet

    Top up and pay per token.

Get an API key Compare speeds

04 / ogbon‑1

Our first model.

ogbon‑1 is in development. We will publish the model card, benchmarks and licence when we release it.

ogbon‑1

In development

Built by
The Ogbon lab, Lagos and London
Designed with
The Ogbon serving runtime
Available on
Ogbon Platform
Model card
At release
Release date
To be announced

We’ll email you when it’s ready.

Get ogbon‑1 updates

05 / Research

Latest from the lab.

Every experiment we run, with what worked, what didn’t, and the numbers behind both.

  1. E004 · In progress

    Hybrid serving of large MoE models

    Experts in system memory, everything else on one GPU. Cheaper per token than GPUs sized to fit them?

  2. E003 · Partial result

    Why AMX bf16 scales poorly across cores

    13.5 TFLOP/s at 8 threads, 5.0 at 16. Noise, kernel switching and power limits ruled out.

  3. E002 · Result

    CPU‑only language‑model serving

    7 to 20 times more per token than a loaded H100. Rejected, so we moved to hybrid serving.

All research

06 / Join the lab

Build it with us.

We are hiring a founding team of engineers and researchers. We hire in Lagos, in London and remotely.

See open roles Meet the team

Try Ogbon today.

Use the platform now. We’ll tell you when ogbon‑1 is ready.

Try the platform Talk to the lab

OGBON.COM

Applied research for fast, affordable AI.

Email us