RingoLLM-v5.1-MuseGlimmer-30B-GGUF

RingoLLM v5.1 (Muse Glimmer 30B)

A LoRA fine-tune of Meta Muse Glimmer 30B specialized for Windows endpoint management: DSC v3, PowerShell 7, and Microsoft Intune.

Narrow domain fine-tunes in this space usually destroy tool calling. This one does not — it matches a dedicated domain fine-tune on accuracy and keeps the base model's full agentic tool-use intact.

Which model should you use

v5.1 refuses legacy DSC by design — see the scope section below.

Benchmarks

Private 25-question rubric covering DSC v3, PowerShell, Intune, legacy DSC, and out-of-scope refusal. Deterministic scoring (must_contain / must_not_contain / regex), identical across all models, temperature 0.3.

Because v5.1 deliberately refuses legacy DSC, the headline comparison excludes that category — scoring a model on work it is designed to decline is not meaningful. Legacy is reported separately below for transparency.

Core capability (23 questions, legacy DSC excluded)

v5.1 is statistically tied with the dedicated domain fine-tune (within 1 point of 131) while also calling tools correctly. The prior fine-tune scores 1/5 on tool selection — and that single point is an abstention case it passes only because it never calls a tool at all.

By domain

Fine-tuning raised DSC v3 from 56% to 89% and Intune from 75% to 88%, without any loss of tool-calling ability.

Scope limit: legacy DSC 1.1

Training on a DSC v3 corpus caused catastrophic forgetting of legacy DSC 1.1. The model scores 39% there, below both the base model (74%) and the prior fine-tune (70%).

Rather than answer unreliably, the shipped chat template instructs the model to refuse legacy DSC requests and offer the DSC v3 equivalent instead. Converting legacy DSC into DSC v3 remains in scope; producing legacy DSC is not.

This is an intentional trade, not a defect being hidden: for legacy DSC work, use v4.2.

Files

Usage (llama.cpp)

llama-server \
  -m RingoLLM-v51-museglimmer-30B-Q4_K_M.gguf \
  --mmproj mmproj-Muse-Glimmer-30B-Q4_K_M.gguf \
  --jinja --chat-template-file ringollm-v51.jinja \
  -ngl 999 -fa on -c 65536 -np 1 \
  --chat-template-kwargs '{"reasoning_strength":"low"}' \
  --temp 1.0 --top-p 0.95 --top-k 64

--chat-template-file is what enforces the legacy-DSC refusal. Without it you get the stock Muse Glimmer template and the model will attempt legacy DSC answers.

Behaviour notes

• It always reasons. Muse Glimmer emits a reasoning channel that cannot be disabled; reasoning_strength (low/medium/high/xhigh) only changes how much. Reasoning is billed against max_tokens, so a small budget can return empty content with finish_reason: length. Allow at least a few hundred tokens.
• Reasoning is returned separately as reasoning_content; content stays clean.
• The base model's dFlash speculative drafter does not transfer. Enabling it against this fine-tune measured slower than no drafter at all. Run without it.

Training

Training data retained reasoning traces deliberately: because the model always reasons at inference, training on reasoning-free data would have created a train/serve mismatch.

The vision tower was intended to be frozen. PEFT suffix-matches bare module names, so vision adapters were attached anyway — but they received no gradient (text-only data never routes through the tower), remained at zero initialization, and were stripped before merging. They are mathematically identity and do not affect outputs.

License

Apache 2.0, inherited from the base model. The training corpus is private and not included.

Attribution

Built on Muse Glimmer 30B by Meta.