What is Modular?
One inference stack that runs models across GPUs, CPUs and ASICs.
Who is Modular best for?
Modular is best for teams that serve AI models in production across NVIDIA, AMD and other chips.
Serving open models at scale
Shared endpoints run models such as DeepSeek, GLM, Kimi and Qwen per token behind an OpenAI-compatible API, so switching from another provider is a one-line change.
Teams on mixed hardware
The same model and code run on NVIDIA, AMD, Google TPU, AWS Trainium, Qualcomm and Apple chips, which avoids being tied to one GPU vendor.
Companies with strict data rules
Bring Your Own Cloud keeps inference inside your own VPC so data never leaves it, and Modular holds SOC 2 Type 2 certification.
Engineers who write GPU kernels
Mojo lets you write or extend kernels for new model architectures, and the self-hosted MAX container costs nothing to run.
How much does Modular cost?
Modular's Self-Hosted edition is free forever, while its hosted Modular Cloud charges per token on shared endpoints, from $0.10 per million input tokens, and per minute for dedicated GPUs.
- Open-source MAX and Mojo in a container on your own hardware
- No usage fees, under the Modular Community License
- Community support on Discord and GitHub
- Shared endpoints billed per token, free to test first
- GPT OSS 120B at $0.10 in and $0.50 out; DeepSeek V4 at $1.74 in and $3.48 out per million tokens
- OpenAI-compatible API, dedicated GPU endpoints billed per minute
Rates on Modular Cloud differ by model, up to $1.74 per million input tokens (DeepSeek V4) and $4.40 per million output tokens (GLM 5.2 and 5.3). Dedicated endpoints are billed per minute of reserved GPU time, so idle time counts, and the page gives no public rate for them. Bring Your Own Cloud (Modular's stack inside your own VPC, billed per minute) and Enterprise are priced through sales, with volume and committed-use discounts available. Full pricing on modular.com. We checked these prices on .
How do readers rate Modular?
No one has rated Modular yet. Be the first.
Sign in to rate. One rating per account.
Is Modular open source?
Yes. Modular's code is public under the Apache 2.0 WITH LLVM-exception licence, so you can run it on your own server.
On GitHub
30,134stars
- Last commit
- 13 Nov
- Repository age
- 3 years
- Latest release
- 22 Oct
- Licence
- Apache 2.0 WITH LLVM-exception
- Self-hosted
- Yes
- Repository
- modular/modular
What are the best Modular alternatives?
Modular is an open source alternative to Replicate and Together AI. Ollama does the same job and is open source too.
Questions about Modular
What is Modular?
One inference stack that runs models across GPUs, CPUs and ASICs.
Who is Modular for?
Modular is best for teams that serve AI models in production across NVIDIA, AMD and other chips.
Is Modular free?
Yes. Modular has a free Self-Hosted plan: open-source MAX and Mojo in a container on your own hardware.
How much does Modular cost?
Modular's paid plans are Modular Cloud at From $0.10 per million input tokens.
Is Modular open source?
Yes. Modular's code is public under the Apache 2.0 WITH LLVM-exception licence, so you can run it on your own server.
What are the best Modular alternatives?
The closest alternatives on Recaply are Replicate, Together AI and Ollama.
Where can I try Modular?
On its official website, modular.com.
Does Modular charge for idle GPU time?
Only on dedicated capacity. Modular Cloud's shared endpoints bill per token used, but dedicated endpoints and Bring Your Own Cloud bill per minute of reserved GPU time, so you pay while it sits idle. The Self-Hosted edition has no usage fees at all.
Can I run Modular on my own hardware?
Yes. Modular's Self-Hosted Community Edition is free and open source, installs through Docker, pip, uv, Pixi or Conda, and runs on NVIDIA, AMD, Qualcomm, TPU, Trainium, Intel, ARM or Apple silicon. Support comes from the community on Discord and GitHub.
Modal
Salad
Tensorfuse
fal.ai
Prime Intellect
CLORE.AI