Muse Glimmer 30B Generates a Complete FastAPI Project in 2 Minutes

A setup connecting Muse Glimmer to the llama.cpp server and the Pi agent allows for coding and testing a project entirely locally, with acceleration provided by DFlash. Measurements report around 46 tokens/s at startup and up to 127 tokens/s on longer tasks. In a FastAPI scenario, a complete project was generated in about 2 minutes. However, limitations remain according to tests, particularly against Qwen3.8-27B for HTML games and applications.

Local tests show 46 to 127 tokens/s and a build in 2 minutes

Local tests measured approximately 46 tokens per second during initial uses, and up to about 127 tokens per second on longer coding tasks. In a complete scenario, a task management API with FastAPI was generated in about 2 minutes, after which the agent was tasked with testing the API and providing a final report. According to this feedback, the agent was quick, smoothly transitioned through steps, and corrected certain issues during debugging in just a few seconds. However, the results remain mixed: a generated HTML game did not function correctly, and Qwen3.8-27B performed better for HTML applications and games.

Pi connects to the local server and detects models without models.json

The installation of the Pi coding agent is done using a single script, followed by the addition of the pi-llama extension from Hugging Face, with a terminal restart recommended. The extension automatically establishes a connection to the local API at http://localhost:8080/v1 and identifies the models offered by llama.cpp without requiring manual configuration of models.json. An empty project can be initialized in /workspace/glimmer-test, and Muse Glimmer appears as an available model through Pi's llama-cpp provider. The Web interface of llama.cpp remains accessible at http://localhost:8080/ for interactive testing.

Download GGUF and start llama.cpp with DFlash on GPU

The main model and the DFlash writer are retrieved from Hugging Face and stored in /workspace/muse-glimmer. The main file muse-glimmer-30B-kquant-17gb.gguf weighs 16.8 GB, and the writer dflash-kquant.gguf is 1.63 GB. The Hugging Face CLI is installed via a script before downloading. llama.cpp is cloned from the official repository, compiled in Release with the CUDA option, and a symbolic link to llama-server is added in /root/.local/bin. The service is launched using speculative decoding DFlash (spec-type draft-dflash), with a spec-draft-n-max set to 15, spec-draft-ngl to all, a context of 16384, and host 0.0.0.0. The server loads the main model and the DFlash writer onto the GPU, with speculative decoding used to accelerate generation.

Announced comparisons and future prospects

In the local ecosystem, Muse Glimmer is compared to 27B class models like those from Qwen, with feedback indicating a frequent advantage for local coding and agentic flows. The described setup was deemed simple, despite some imperfections still present. Improvements are anticipated for Muse Glimmer, llama.cpp, DFlash, and associated tools, aiming for better results, higher speeds, and enhanced agentic performance locally. It is also suggested to try these models on RTX 3090, 4090, or 5090 cards, while the experience with local models is described as approaching that of systems like GLM-5.2.