Updated
Updated · KDnuggets · Aug 21
Muse Glimmer Runs Locally on RTX 3090, Reaching 127 Tokens/Second for Agentic Coding
Updated
Updated · KDnuggets · Aug 21

Muse Glimmer Runs Locally on RTX 3090, Reaching 127 Tokens/Second for Agentic Coding

3 articles · Updated · KDnuggets · Aug 21

Summary

  • Muse Glimmer was deployed locally through llama.cpp with DFlash and Pi, letting the 30B coding model build, test and debug projects from a terminal on an RTX 3090.
  • A 16.8 GB main model and 1.63 GB drafter used speculative decoding to lift throughput from about 46 tokens per second in initial tests to roughly 127 during longer coding runs.
  • In one end-to-end trial, the model built a FastAPI task-management API with SQLite and pytest in about 2 minutes, then identified and fixed errors during debugging.
  • Results were uneven: Muse Glimmer impressed on multi-step local coding workflows, but the tester said Qwen3.8-27B still produced better working HTML apps and games.
  • The setup underscores a broader push toward private local AI coding, with the report arguing open models are narrowing the gap with proprietary systems on high-end consumer GPUs.

Insights

Will speculative decoding finally make expensive cloud-based coding assistants completely obsolete for everyday developers with consumer GPUs?
Is investing in power-hungry hardware like the RTX 3090 truly justified when cloud AI writes code for pennies?
Can a private AI coding agent with autonomous shell access become a massive security threat to your own local machine?