Over the last year, we've been obsessed with one problem: making AI inference faster and more efficient. As we worked with developers building AI products, we kept seeing the same tradeoff:
Use closed-source AI APIs and give up privacy and control. or Run open-weight models yourself and spend time managing infrastructure.
We built Zro so developers don't have to choose.
With Zro, you get a fast, OpenAI-compatible API for open-weight models with:
🌏 Multi-region hosted inference 🔒 Zero data retention & zero training on your prompts 🏢 Optional on-prem deployments ⚡ A serving stack optimized for coding agents and long-context workloads using MoonMath's in-house inference technology. 🛠️ Built for easy setup with the most popular AI coding agents
We're launching early because we want feedback from developers building real AI products. We'd love to know:
Which models should we add next? What features are missing from your current inference provider? What would make you switch?
🎁 Product Hunt launch offer
Use code PRODUCTHUNT to get 1 month of Zro Pro free. Limited to the first 100 users.
We'll be here throughout the day to answer every question. Thanks for checking out Zro, we're excited to hear what you think! 🚀
Report
The tradeoff between closed APIs and self-hosting open models is very real. most small teams want more privacy and control, but not the operational burden of running inference infrastructure themselves.
The OpenAI-compatible API, zero retention, and optional on-prem path make Zro feel practical for teams that want to start simple and keep more control as they grow. Curious how you handle routing across regions for long-context coding workloads without hurting cache efficiency or latency consistency?
@andrasczeizel We are not routing your request across region and that's how it wouldn't hurt the cache at all. Before setting up your session with API key you will be able to select which region you want to use.
Right now, all live machines are located in the EU, but US options will be available soon!
Report
@emirsoyturk I see, thanks for the reply. Good luck with the launch & vercel day!
Report
Congrats on the launch! Keeping it OpenAI-compatible is a massive smart move. A lot of teams want to switch to open-weight models for privacy but dread rewriting their entire agent orchestration layer or dealing with completely different API schemas. How has the drop-in replacement experience been for early testers? Are there any specific edge cases with function calling or streaming where the compatibility layer behaves differently than native OpenAI?
@franz_briones Thanks for your comment Franz. It's really difficult problem and ZRO has built-in support for many popular harnesses like Claude Code, Codex, Pi, OpenCode. You can see all of the ones we officially support here: https://zro.moonmath.ai/integrations. We are also adding new harnesses like Kilo Code, Grok Build, Command Code.
Which open-weight models would you like to see? Is there any tool that you're using and ZRO doesn't list it as supported?
Report
"Your code shouldn't be someone else's training data" is a strong line. Which open models are you running under the hood, and is EU hosting the default or a paid tier? Congrats on the launch
@alex_tomilin EU hosting is the default and we are working on adding new regions (e.g US). Right now GLM 5.2 and Minimax M3 are accesible via ZRO and we are testing/optimizing new models.
Do you have any models that you would like to see added?
Report
Private inference is the box teams check last, usually right after their proprietary code turns up somewhere it should not. The part I would pressure-test is latency ten tool calls deep in a real agent loop, not on a single prompt. That is where most self-hosted setups fall over.
@shivangit26 Exactly, that is the reason we are optimizing our inference for long-context agentic workflows!
Let us know if there is any model you would like to see.
Report
Streaming support for token-by-token responses would be huge for chat use cases, even if it means slightly more work on the caching layer. Right now waiting for full completions before anything renders feels dated for an inference API in 2026.
@paige_lauren1 Fair point. Privacy has always been core to what we build at MoonMath and Ingonyama. We're actively working toward stronger transparency around these guarantees.
Zro
👋 Hi Product Hunt!
We're MoonMath, the team behind Zro.
Over the last year, we've been obsessed with one problem: making AI inference faster and more efficient. As we worked with developers building AI products, we kept seeing the same tradeoff:
Use closed-source AI APIs and give up privacy and control.
or
Run open-weight models yourself and spend time managing infrastructure.
We built Zro so developers don't have to choose.
With Zro, you get a fast, OpenAI-compatible API for open-weight models with:
🌏 Multi-region hosted inference
🔒 Zero data retention & zero training on your prompts
🏢 Optional on-prem deployments
⚡ A serving stack optimized for coding agents and long-context workloads using MoonMath's in-house inference technology.
🛠️ Built for easy setup with the most popular AI coding agents
We're launching early because we want feedback from developers building real AI products. We'd love to know:
Which models should we add next?
What features are missing from your current inference provider?
What would make you switch?
🎁 Product Hunt launch offer
Use code PRODUCTHUNT to get 1 month of Zro Pro free.
Limited to the first 100 users.
We'll be here throughout the day to answer every question. Thanks for checking out Zro, we're excited to hear what you think! 🚀
The tradeoff between closed APIs and self-hosting open models is very real. most small teams want more privacy and control, but not the operational burden of running inference infrastructure themselves.
The OpenAI-compatible API, zero retention, and optional on-prem path make Zro feel practical for teams that want to start simple and keep more control as they grow. Curious how you handle routing across regions for long-context coding workloads without hurting cache efficiency or latency consistency?
Zro
@andrasczeizel We are not routing your request across region and that's how it wouldn't hurt the cache at all. Before setting up your session with API key you will be able to select which region you want to use.
Right now, all live machines are located in the EU, but US options will be available soon!
@emirsoyturk I see, thanks for the reply. Good luck with the launch & vercel day!
Congrats on the launch! Keeping it OpenAI-compatible is a massive smart move. A lot of teams want to switch to open-weight models for privacy but dread rewriting their entire agent orchestration layer or dealing with completely different API schemas. How has the drop-in replacement experience been for early testers? Are there any specific edge cases with function calling or streaming where the compatibility layer behaves differently than native OpenAI?
Zro
@franz_briones Thanks for your comment Franz. It's really difficult problem and ZRO has built-in support for many popular harnesses like Claude Code, Codex, Pi, OpenCode. You can see all of the ones we officially support here: https://zro.moonmath.ai/integrations. We are also adding new harnesses like Kilo Code, Grok Build, Command Code.
Which open-weight models would you like to see?
Is there any tool that you're using and ZRO doesn't list it as supported?
"Your code shouldn't be someone else's training data" is a strong line. Which open models are you running under the hood, and is EU hosting the default or a paid tier? Congrats on the launch
Zro
@alex_tomilin EU hosting is the default and we are working on adding new regions (e.g US). Right now GLM 5.2 and Minimax M3 are accesible via ZRO and we are testing/optimizing new models.
Do you have any models that you would like to see added?
Private inference is the box teams check last, usually right after their proprietary code turns up somewhere it should not. The part I would pressure-test is latency ten tool calls deep in a real agent loop, not on a single prompt. That is where most self-hosted setups fall over.
Zro
@shivangit26 Exactly, that is the reason we are optimizing our inference for long-context agentic workflows!
Let us know if there is any model you would like to see.
Streaming support for token-by-token responses would be huge for chat use cases, even if it means slightly more work on the caching layer. Right now waiting for full completions before anything renders feels dated for an inference API in 2026.
Zro
@frat8d8j Totally agree. ZRO already supports streaming through the OpenAI-compatible API.
If you saw a buffered response, let us know which client or integration you used!
"zero retention claims are cheap, anyone can audit this?"
Zro
@paige_lauren1 Fair point. Privacy has always been core to what we build at MoonMath and Ingonyama. We're actively working toward stronger transparency around these guarantees.