The Runpod Blog

Why your balance can drop even when nothing succeeds, how to find where the money went, and how to keep it from happening again.

Less of the work around your compute is yours to maintain this quarter, from one API for the whole platform to one account for your whole company. Plus the deadlines for v1 and GraphQL users.

Flash now ships a first-class Nix flake. Run the CLI with nix run github:runpod/flash, get a pinned dev shell, and send GPU work to Runpod Serverless.

You can teach an open-weight LLM to text like a specific persona: their slang, their reply length, their humor, even how they write in a roleplay scene, creative writing exercise, or chatbot customer interactions. All it takes is some chat logs (synthetic or real), a few hours of data prep, and a single GPU pod. This guide walks through the whole process on Runpod with Axolotl, from raw messages to a model you can chat with in SillyTavern, KoboldCpp, or text-generation-webui, using two base models: a 27B Qwen model and the 128B Mistral Medium 3.5.‍ Everything here was tested end to end, and every error we hit along the way is in the Troubleshooting section, so you shouldn't get stuck where we did.

Attach a Global Volume to a Runpod Serverless endpoint and your workers start in any data center with your GPU types. Store model weights once, read anywhere.

Deploy OpenJev, a 27B decision model that returns structured answers and confidence scores, on a Runpod Serverless load balancer endpoint, then send your first request with cURL and Python.

The State of AI Compute shows what teams actually deploy: agents are 8% of users but drive 24% of revenue, a new model version can take over its family's workloads in about a month, and lower precision is becoming a production strategy for large models.