Muse Glimmer 30B now runs on Xyntetik Runner, if you need an option to Llama/ollama for built-in schema support.
Muse Glimmer 30B now runs on Xyntetik Runner, my from-scratch local-LLM inference engine in C. Support landed in the current build: https://github.com/Joakimpalm-Zen/xyntetik-runner/releases/tag/v0.1.15-alpha.
Get started:
runner -m muse-glimmer-30B-kquant-17gb.gguf --serve
The mechanism here is not ordinary JSON Schema support, it is forced-truncation recovery: once a document starts, Runner emits the smallest schema-legal ending when the token budget expires, so the arguments still parse. On local models, where context is tight and generation is slow, that is the difference between an agent loop that finishes and one that crashes.
Repo: github.com/Joakimpalm-Zen/xyntetik-runner
Update: Muse's native tool calling now works, and it fits this engine unusually well.
Muse does not ask for tools in JSON. It has its own format, atem, where a tool call starts by naming the tool and then lists the arguments one by one.
That order is the whole trick. Runner constrains generation token by token with a schema compiler, and the hardest part of constraining is knowing early which tool the model is calling. atem tells you in the first line. So Muse's own format slots straight into the machinery Runner already had: every tool call comes out well-formed, aimed at a declared tool, with arguments matching the schema. No parser guessing, no retries.
The truncation trick carries over too. If the token budget runs out mid-call, Runner closes the atem block legally, so your agent still gets a call it can execute instead of half a document.
And the model is never forced to speak a foreign format. It writes the exact syntax it was trained on, its own protocol tokens and all, which keeps it on-distribution and the output quality where it should be.
Shipped in v0.1.16-alpha: https://github.com/Joakimpalm-Zen/xyntetik-runner/releases/tag/v0.1.16-alpha