Open Model Hosting & GPUs
Run open-weight or custom models in production without buying hardware — either as a ready API or on GPUs you rent.
Trends
Modal gained the most ground in package downloads: 12.6× against the median npm package over a year.
Public counts each month: pull requests on GitHub, package downloads, stars, Product Hunt followers. They show direction, not market share.
The real choice
Per-token API or your own code on rented GPUs. Serverless model APIs host popular open models behind an OpenAI-compatible endpoint, so you pay per token and run nothing. Renting GPU compute lets you run any model or pipeline you like, but you package it, handle cold starts and pay for time, busy or idle.
Pick by situation
Tap the ones that are you — the tools that fit light up below.- The open model you need is already served by a host and you want to pay per token
Together AI
Fireworks AI
Hugging Face - Your workload is your own Python code — a pipeline, batch job or custom model — that should scale to zero
Modal - You want a raw GPU machine for experiments, training or a long-running server
Runpod - One model is core to the product and needs steady latency under real traffic
Baseten
Hugging Face - You run the GPUs yourself and need an OpenAI-compatible server for many concurrent users
vLLM
Compare the tools
Grouped by the side of the choice they answer, not ranked. Open a row for when to use it, the trade-off and what makers say.
Together AIplatformCalling many popular open models through one OpenAI-compatible API, with fine-tuning on the same platform.


12+33 open source1M/wk3.8× vs npm
Use it whenYou want open models without managing any infrastructure.
Trade-offYou're limited to the models they host unless you pay for dedicated capacity.
Used by AgentR, Hedy AI, TextJam, Arcee, Ariadna and 7 more · in 33 open-source projects.
Together AI makes a wide variety of open source models accessible, at an industry leading cost. It lets us provide access to modules like Llama and a privacy-friendly versions of DeepSeek.
Fireworks AIplatformLow-latency serving of open models, including your own fine-tuned versions.


5+16 open source561.1k/wk+177% vs npm
ModalplatformRunning your own Python code — inference, batch jobs, sandboxes — on GPUs defined in code, scaling to zero between calls.


60+22 open source671.8k/wk12.6× vs npm
Use it whenYour workload is more than a single model call, or you need a model no API serves.
Trade-offCold starts on large models, and your code is shaped around Modal's decorators.
Used by Almanac, Clusy, DashGPT, Deeptrue, Gan.AI and 55 more · in 22 open-source projects.
Modal helps us spin up Python workloads with speed and control. Great for agents that need to run fast, ephemeral jobs. A dream for short-lived compute without DevOps overhead.
RunpodplatformRenting GPU machines by the hour or running serverless GPU workers, with control over the hardware.


19+3 open source144.4k/wk
Use it whenYou want a raw GPU box for experiments, training or a long-running server.
Trade-offMore hands-on setup and operations than a managed platform.
Used by Glyp, Hero, TensorPool, Virtualstagingai, Autonomous and 14 more · in 3 open-source projects.
Runpod powers our AI video generation behind the scenes. Affordable GPU compute with great reliability and speed.
Hugging FaceplatformDeploying almost any model from the Hugging Face Hub to a dedicated endpoint, or calling popular models through Inference Providers, which routes to partner hosts under one API.


77+191 open source387.7k/wk−46% vs npm
Use it whenThe model you need lives on the Hub — including your own fine-tune — and you want it served without packaging it yourself.
Trade-offDedicated endpoints bill per hour while running; Inference Providers only covers models the partner hosts serve, at their prices.
Used by Cotypist, Free AI Video Editor OpenCutAI, IFTTT, Inquio, KushoAI and 72 more · in 191 open-source projects.
Our in-browser AI features use Hugging Face Transformers.js. Running ML models directly in the browser without a backend call that's the magic Hugging Face enables for local-first AI apps.
Basetenplatform · self-hostableDeploying a specific model as a production endpoint with autoscaling and monitoring.


23+2 open source80.3k/wk
Use it whenOne model is core to your product and needs steady latency under real traffic.
Trade-offBuilt for production workloads; more setup than calling a shared serverless API.
Used by Not Diamond, Cekura, SpeechifyAI, You.com, AudioForgeAI and 18 more · in 2 open-source projects.
Baseten plays a central role in our AI infrastructure for hosting and serving our model recommendation engine.
vLLMlibrary · open sourceServing open-weight models on your own or rented GPUs behind an OpenAI-compatible API, with high throughput for many concurrent requests.43 open source446.5k/wk
Use it whenYou've outgrown per-token APIs on volume, or need a model or fine-tune no host serves, and have GPUs to run it on.
Trade-offYou own the GPUs, scaling, upgrades and on-call; tuning memory and batching settings takes real effort.
No maker's product we track shows it for this yet · in 43 open-source projects.
Before you choose
If the model you need is already served as an API, use that — it's the least work and scales to zero. Rent GPUs only for custom models, fine-tunes a host won't serve, or non-LLM pipelines. Check cold-start times with real traffic before promising users response times.
- Leaving a dedicated endpoint or rented GPU running over a weekend of no traffic; hourly billing doesn't care whether anyone called it.
- Promising users sub-second responses before measuring cold starts on a large model under real traffic.
- Renting GPUs to serve a popular open model that a serverless API already hosts for less than the idle time would cost.
Other options
Real choices most makers here won't need to weigh.
Decided alongside
What the 159 makers' products here chose for their other decisions.