Basalt hits 75 tok/s on a 5060 Ti: local LLMs on modest hardware
ironicstatistic · reddit · 2026-10-15
A Reddit user shares results with Basalt, a fork of Strata optimized for Blackwell GPUs by jesdga95: on a system with just an RTX 5060 Ti 16GB and 32GB RAM, it reaches 75 tok/s generation and 1200 tok/s prefill at 128k context using a Q2 model with Q4 KV cache — usable as a dedicated coding agent.
The developer even reached out to help him tune it directly in the comments; the user ultimately had DeepSeek add a build flag to allocate more RAM to make it work. Open-sourced on GitHub.
More from coding & agent
- Figma Launches Build by Design Workflows to Bridge Specs, Code Layers and Coding Agents — abhi9u · 2026-10-15
- Figma Make adds Specs and Payload integration for prototype-to-ship workflows — abhi9u · 2026-10-15
- Figma Specs turns canvas selections into coding-agent-ready specs — abhi9u · 2026-10-15
- Figma Make integrates open-source CMS Payload for instant backends — abhi9u · 2026-10-15
- Figma adds Code layers, Specs and design system support to court developers — abhi9u · 2026-10-15
- Weaviate Query Agent ships a native MCP server with one-URL setup — CShorten30 · 2026-10-15