Koyeb

A French serverless platform that runs a Git repository or container image as autoscaling microVMs on its own bare-metal servers, on CPUs or GPUs and down to zero, which agreed in February 2026 to join Mistral AI.

What it is: Platform as a service (PaaS)
Used for: MicroVMs
Type: Commercial product
Made by: Koyeb SAS
License: Proprietary
Facts checked: 13 October 2026

What it is

You give Koyeb a Git repository, which it builds with buildpacks or your Dockerfile, or a pre-built container image, through the dashboard, the koyeb CLI, the API, Terraform or Pulumi. Each running copy is an Instance: a microVM on bare-metal servers that Koyeb operates in six core regions (Frankfurt, Washington, D.C., Singapore, Tokyo, Paris and AWS us-east-1, with San Francisco in preview). An edge network sends each user to the nearest region where the app runs. Instances are billed per second: a Standard nano (0.25 vCPU, 256 MB) costs $2.68 a month, an L4 GPU $0.70 an hour, an H100 $2.50 an hour and eight H200s $24 an hour. The autoscaler adds Instances on requests per second, concurrent requests, p95 response time, CPU or memory.

Koyeb first ran user apps on Kubernetes, then in 2021 moved them to its own stack, where Nomad schedules each Instance as a microVM. Its own Envoy load balancers, configured from a Consul registry in each region, route every request to the nearest region running the app. It began with Firecracker, but Firecracker can’t pass GPUs through to a VM, so it switched to Cloud Hypervisor behind Kata Containers. That switch made snapshots possible: with Scale-to-Zero, an idle service either enters Deep Sleep (a new VM on the next request, a 1 to 5 second cold start) or Light Sleep, where the VM is paused to a snapshot and resumes in about 200 ms. Sandboxes, in public preview since November 2025, use the same microVMs to run AI-generated or untrusted code through Python and JavaScript SDKs.

On February 17, 2026 Koyeb announced a definitive agreement to join Mistral AI, subject to closing conditions. Its whole team is moving to Mistral Compute, which the platform is to become a core part of; meanwhile it keeps operating and billing as before. New sign-ups must choose a paid plan (Pro at $29 a month with $10 of compute included, Scale or Enterprise); the pay-as-you-go Starter plan, with its one free Instance, stays only for existing organizations. In October 2026 Koyeb’s site still says it “is joining” Mistral AI, its public changelog has had no entry since February 27, 2026, and its CLI and SDKs still get releases.

Why teams pick it

Git push or a container image becomes a load-balanced, autoscaling service on CPUs or GPUs (from an RTX 4000 SFF Ada at $0.50 an hour to 8x H200), billed per second
Scale-to-Zero for CPU and GPU services, with Light Sleep snapshots that wake a CPU Instance in about 200 ms
Every Instance is a microVM, a stronger boundary between customers than containers sharing a kernel, and the same machinery runs agent sandboxes
One service can run in several regions at once, behind Koyeb’s global load balancers and private network

Watch out for

Its future is tied to Mistral AI: the team now works on Mistral Compute, the platform is to be folded into it, and new users can no longer start free
Scale-to-Zero and Light Sleep are still in public preview; only services exposed to the internet can scale to zero, HTTP/2 requests can’t wake a sleeping service, and Deep Sleep cold starts take 1 to 5 seconds
Few regions: six core regions, Eco Instances only in Washington, D.C., Frankfurt and Singapore, managed PostgreSQL only in those three, and GPUs only in some regions
Proprietary: the engine isn’t open source, so leaving means redeploying elsewhere

Alternatives

What all platform as a service (paas) have in common →Compare side by side →

HerokuCommercialThe platform that defined PaaS: git push your code, and Heroku builds it with a buildpack and runs it on managed containers called dynos behind its own HTTP router.

RenderCommercialRender builds and deploys web services, background workers, cron jobs and managed Postgres and Key Value stores from your Git repo, all described in one render.yaml file if you like.

RailwayCommercialA usage-billed platform that builds a repository or runs an image as services on a visual project canvas, with databases, volumes, private networking and per-pull-request environments, on hardware Railway runs itself.

Fly.ioCommercialFly.io turns your Docker image into fast-booting Firecracker microVMs on its own servers in regions around the world, and routes each user to the nearest one.

NetlifyCommercialA web platform that builds your site from Git, publishes each build as an immutable, atomic deploy on its CDN, and runs server code beside it as functions, edge functions and a branchable Postgres database.

How it fits with other technologies

What Koyeb uses

Deploy an image to two regions with the Koyeb CLI, scaling between 0 and 5 Instances · bash

koyeb apps init my-api \
  --docker ghcr.io/acme/api:1.4 \
  --instance-type small \
  --regions fra --regions was \
  --ports 8000:http --routes /:8000 \
  --min-scale 0 --max-scale 5 \
  --autoscaling-requests-per-second 100 \
  --light-sleep-delay 5m --deep-sleep-delay 1h

See it in a real stack

No architecture uses Koyeb yet. Drop it into the playground and wire it up yourself.Open playground →

Where it came from2020–2026

Yann Léger, Édouard Bonlieu and Bastien Chatelard spent the 2010s building and running cloud service providers. In 2019 the three started Koyeb, betting that serverless and multi-cloud would shape the next decade, and in October 2020 they raised a €1.4 million pre-seed round led by ISAI, with angels such as Datadog co-founder Alexis Lê-Quôc and Packet co-founder Zachary Smith.

The first version ran on Kubernetes. Within months the team moved user workloads to a stack of its own on bare metal: Nomad for orchestration, Firecracker microVMs for isolation and Kuma with Envoy for networking, keeping Kubernetes only for the control plane. After a private preview with more than 10,000 developers, the platform opened to everyone in May 2022, raised a $7 million seed round led by Serena in November 2023, and added GPUs in June 2024, which led it to replace Firecracker with Cloud Hypervisor.

As AI workloads grew, Koyeb added Light Sleep in August 2025, sandboxes for AI agents in November 2025 and newer GPUs (RTX Pro 6000, H200, B200) in January 2026. On February 17, 2026 the founders announced that Koyeb would join Mistral AI, which wanted infrastructure for its own cloud, Mistral Compute.

Why it’s built this way

Nomad instead of Kubernetes for user workloads: Kubernetes was hard to extend, needed a full cluster per zone for multi-zone deployments, and its kubelet used 10% to 25% of each hypervisor’s RAM. Nomad took a custom Firecracker driver easily, was multi-zone out of the box and needed about 100 MB.
A microVM per Instance: Koyeb runs many customers on the same servers. It tried gVisor first but found the performance disappointing, and running Firecracker on Kubernetes was experimental, so it built around microVMs directly.
Its own stack on bare metal, not a hyperscaler’s primitives: Building on another cloud’s function and container services would have meant inheriting their limits; owning the core technology let Koyeb control performance and cost.
Cloud Hypervisor behind Kata Containers: Firecracker lacks PCI passthrough, so it couldn’t run GPU Instances. Kata lets Koyeb swap virtual machine monitors, and Cloud Hypervisor’s snapshots made 200 ms Light Sleep possible.

Timeline

2020
€1.4M pre-seed round led by ISAI, in October
2021
User workloads move from Kubernetes to Nomad, Firecracker and Kuma on bare metal
2022
Public preview in May, after a private preview with over 10,000 developers
2023
$7M seed round led by Serena, in November
2024
Serverless GPUs in public preview, in June
2025
Light Sleep wakes services in about 200 ms (August); Sandboxes in public preview (November)
2026
Definitive agreement to join Mistral AI, February 17

Sources15 · facts checked 13 October 2026

Improve this page on GitHub