SASVAAI/muse-glimmer-30b-kernelgen

VISIONPricing:Input $1.2 / Cached $0.04 / Output $4.4Concurrent Unit Cost:2Model Size:30BQuant:FP8Context Size:128kPublished:Sep 14, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

SASVAAI/muse-glimmer-30b-kernelgen is a 30 billion parameter merged fine-tune of Meta's Muse-Glimmer-30B model, developed by SASVA AI Model Cognition Labs (MCL) Team. This model specializes in generating Triton GPU kernels from PyTorch nn.Module specifications, providing self-contained Python host code and `@triton.jit` kernels. It is optimized for developers porting PyTorch code to high-performance Triton implementations, trained with QLoRA on a context length of 8,192 tokens.

Loading preview...

Model Overview

SASVAAI/muse-glimmer-30b-kernelgen is a 30 billion parameter model developed by SASVA AI Model Cognition Labs (MCL) Team, specifically engineered for Triton GPU kernel generation. It is a merged fine-tune of the meta-models/Muse-Glimmer-30B base model, with its LoRA adapter fused into the base weights, making it loadable as a standard transformers checkpoint without PEFT dependencies.

Key Capabilities

  • PyTorch to Triton Conversion: Translates PyTorch nn.Module specifications into complete, self-contained Python files with @triton.jit kernels and host-side launch code.
  • GPU Kernel Engineering: Designed for developers who need to port nn.Module code to hand-written, high-performance Triton kernels.
  • Context Length: While the base model supports 131,072 tokens, this fine-tune was trained at an effective context length of 8,192 tokens.
  • Ollama Support: A Q4_K_M GGUF version is included for local execution with Ollama, offering a 15.8 GiB quantized option.

Intended Use Cases

This model is ideal for:

  • Accelerating PyTorch Workloads: Generating optimized Triton kernels for specific PyTorch operations to improve performance.
  • GPU Programming: Assisting GPU kernel engineers in writing triton.jit code by automating the translation from PyTorch modules.
  • Research and Development: Exploring efficient kernel implementations for custom PyTorch layers or functions.

It's crucial to use the exact prompt format it was trained on, as it is not intended as a general chat model. The model was trained on a dataset including GPUMODE/KernelBook and GPUMODE/kernelbot-data, focusing on permissively licensed and correctly running kernel implementations.

Finetunes
1 models