wtf is repomind-v3?

REPOMIND v3 is an on premises coding assistant meant for companies that cannot send their code to an outside cloud AI service, such as regulated banks or firms with strict data policies. Instead of calling a service like OpenAI or Anthropic, it runs entirely on AMD hardware, using open weight models such as Gemma, GLM, and Qwen3 Coder, either through AMD Developer Cloud, AMD hosted Fireworks AI, or fully on premises on an AMD Instinct MI300X GPU. The system routes questions based on size. Shorter questions, under about 8,000 tokens, go to a smaller Gemma model for a quick first answer. Larger questions get a two step process: Gemma first looks at a short summary of the codebase's files and picks out which ones are actually relevant, then a bigger model reasons deeply over just those selected files instead of the whole codebase, which the README says reduces wasted computing time. There is also an air gapped mode that forces every query to stay on the local AMD hardware and blocks any outside connection entirely, intended for the strictest compliance settings. Beyond answering questions about a codebase, it includes a tool that converts CUDA code, the kind written for NVIDIA GPUs, into ROCm code that runs on AMD GPUs, using a set of mapping rules for memory, streams, events, and kernel launches, and it outputs the change as a diff you can apply directly. It also keeps a tamper evident audit log using cryptographic signing and hash chaining, and supports an API key requirement for its endpoints. To try it, you clone the repository, copy an example environment file and add a Fireworks API key if you have one, then run it with Docker Compose. There is also an offline demo mode that runs the whole agent loop against a mock model with no GPU and no API keys required, useful for trying the interface without any cloud account. This is described as version 3 of the project, building on an earlier version that was benchmarked directly on AMD MI300X hardware.