What's up DEV Community👋 I’m Richard from MangoBoost, and we just launched Mango Inference, an API for serving open-weight models including DeepSeek V4 Pro, GLM-5.3, Kimi K3, and MiniMax M3.
It’s OpenAI-compatible, so switching usually means changing your base_url, API key, and model name, with no other code changes.
A couple of things people usually ask about:
Data retention: Zero data retention is on by default. We don’t store your prompts or outputs. The one exception is a short-lived prefix cache, which makes repeated prompts faster and cheaper and isn’t tied to your account.
Agents and long sessions: that same cache stays alive [for about an hour], so resumed agent sessions can keep hitting it instead of paying full price to resend the whole context.
Where it runs: we run the models on our own hardware, so your requests never go to a third-party model API.
The launch offer: $10 in free credits when you sign up, and 30% off all models through Nov 11.
👉 Pricing and Free Credits:
What we’d really like back is honest feedback: latency on your actual prompts, how it handles agent or tool-calling workloads, rough edges in the docs, anything that bugs you. Results where we lose are just as useful to us.
Questions are welcome in the comments, and I’ll pass the technical ones along to our engineers.
Top comments (2)
Faster Inference service! 🥳🥳
I’ve been curious about GLM 5.3, so I’ll give it a try with this.