The question comes from a user with a GeForce RTX 5070 Ti X3 with 16GB of VRAM and 32GB of DDR5 RAM. His goal is ambitious but consistent with the local AI promise: one agent for daily tasks, one for research, and one for development. He tried Ollama and vLLM, then the Goose, Hermes Agent, and OpenCode frameworks. The declared result is zero: no tool proved useful. Not for lack of trying, but because of a lingering doubt: is it a skills issue, a tooling issue, a model configuration issue, or all of the above?

The answer isn't in the single post, but the described setup is a textbook case for understanding where on-prem AI gets stuck today. With 16GB of VRAM, the margin for keeping model, context, tool calling, and orchestration together is tight. Larger models that handle the multi-step reasoning required by agents simply don't fit without aggressive quantization. And when quantization goes too far, it degrades exactly the capabilities an agent needs: following instructions, planning, using external tools. This isn't a flaw in a single harness; it's a structural variable.

Then there is the orchestration layer. Goose, Hermes Agent, and OpenCode come from different philosophies: some favor integration with local tools, others focus on conversational automation, others on pair programming. But none of the three, according to the account, turned the setup into a reliable assistant. The interesting fact isn't the failure itself, but that the user doesn't know whether the problem is the model, the framework, or the configuration. That's exactly the point: self-hosted isn't a product, it's a construction site. Those coming from cloud, where an API endpoint and a key are enough to start, collide with an environment where every choice has a cost in VRAM, context, and latency.

For those evaluating on-premise deployment, there are trade-offs between control and operational complexity that a single framework cannot solve. The case signals something structural: the race toward local agents is pushing consumer hardware downward, but frameworks haven't yet closed the gap. Who wins today? GPU makers with more memory and teams developing compact models optimized for tool calling. Who loses? The individual developer or small team that believes a mid-range card is enough to replicate the cloud assistant experience. Data sovereignty and TCO, which make self-hosted attractive, come at the cost of a learning curve and an uncertain yield. The request for suggestions, after all, is still open.