SASVAAI/Muse-Glimmer-30B-rego

VISIONPricing:Input $1.2 / Cached $0.04 / Output $4.4Concurrent Unit Cost:2Model Size:30BQuant:FP8Context Size:128kPublished:Oct 13, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

SASVAAI/Muse-Glimmer-30B-rego is a 30B parameter merged fine-tune of Meta's Muse-Glimmer-30B model, developed by SASVA AI Model Cognition Labs (MCL) Team. It is specifically optimized for generating Open Policy Agent (Rego) policies from plain-language compliance controls and target package names, supporting a context length of 131,072 tokens. This model excels at authoring policy-as-code for Terraform, Kubernetes, and cloud resource documents, making it ideal for compliance and infrastructure-as-code automation.

Loading preview...

Muse-Glimmer-30B-rego: Rego Policy Generation

SASVAAI/Muse-Glimmer-30B-rego is a specialized 30 billion parameter language model, fine-tuned from Meta's Muse-Glimmer-30B. Developed by SASVA AI Model Cognition Labs (MCL) Team, this model is designed to generate Open Policy Agent (Rego) policies directly from natural language compliance requirements and specified package names. It operates with a substantial context length of 131,072 tokens, though it was trained at 2,048 tokens.

This model is a merged fine-tune, meaning its LoRA adapter has been fused into the base weights, allowing it to load as a standard transformers checkpoint without PEFT dependencies. It was trained using 4-bit-quantized-base LoRA via TRL SFT, focusing on attention projections only.

Key Capabilities

  • Rego Policy Generation: Translates plain-language compliance controls into complete and correct Rego policies, including package declarations and deny rules.
  • Policy-as-Code Automation: Streamlines the creation of policies for infrastructure-as-code (IaC) tools like Terraform and Kubernetes, and cloud resource documents.
  • Specific Prompt Format: Requires a precise prompt structure, including control requirements, target package, rule, and optional input shape, to ensure accurate output.
  • Multimodal Architecture (Text-focused): While based on a multimodal architecture, the fine-tuning focused exclusively on the language model for text-to-Rego generation; the vision tower remains untouched.

Good for

  • DevOps and Security Engineers: Automating the creation of compliance policies for cloud environments and IaC.
  • Compliance Automation: Generating Rego policies to enforce security and regulatory compliance controls.
  • Policy Development: As a starting point for authoring Rego policies, particularly for Terraform and Kubernetes resource checks.

Note: Generated policies are intended as a starting point and require review and testing. The model's Q4_K_M quantization may produce semantically different policies compared to bf16 or Q8_0 due to fidelity differences, with Q8_0 showing byte-identical output to bf16 on tested prompts.

Finetunes
1 models