mesh-llm pools GPUs for open-model inference

This is more compelling as a coordination layer than as a raw inference engine. If the auto-configured mesh is stable under churn, it could make shared private model hosting feel like an application instead of infrastructure.

• –It supports dense-model splitting, MoE expert sharding, demand-aware rebalancing, and Nostr-based discovery.
• –The OpenAI-compatible endpoint and launcher integrations mean existing tools can point at it with little or no glue.
• –The blackboard layer adds agent collaboration on top of inference, which is a stronger product story than “distributed GPU router” alone.
• –The main risk is reliability: spare capacity is volatile, so retry behavior, node churn, and latency variance will decide whether this feels robust or brittle.