studiocatalyst/atlas-models

Catalyst Atlas models

The model files Catalyst Atlas downloads on first use, each pinned by its SHA-256. They are other authors' models, converted or quantized to run locally; none was trained or fine-tuned here. Each keeps its own licence, below.

File	Model	What was done	Licence	SHA-256
embeddinggemma-2-UD-Q5_K_XL.gguf	EmbeddingGemma 2 by Google DeepMind: its text model	None: an unchanged copy of Unsloth's dynamic Q5_K_XL quantization	Apache 2.0	a9d7a3b7eb421113f05e964e10882676f10fff83811b7876d5bb983db7c7ac41
mmproj-Q5_1.gguf	EmbeddingGemma 2 by Google DeepMind: its picture and sound encoders with their projectors	Quantized to Q5_1 (the linear layers only) from ggml-org's BF16 GGUF	Apache 2.0	cc85c7137a34989ef5fb6fee08725671c936276cff5eb820b93ff02c39528906
omnishotcut_v1.5_fp16.onnx	OmniShotCut v1.5 by the UVA Computer Vision Lab: finds the cuts between shots	Converted to ONNX, in 16-bit floats	MIT	9c39c117eafe6cd6939c6236ccb274076c7c9c4a2927de3ac606246694ac05d3
Qwen3-ASR-1.7B.i1-Q5_K_M.gguf	Qwen3-ASR 1.7B by the Qwen team, Alibaba Cloud: speech recognition	None: an unchanged copy of mradermacher's Q5_K_M with an importance matrix	Apache 2.0	737f4b9cb049361dad7d4625d298b16a0a2cb026268da195b93298329636a304
Qwen3-ASR-1.7B.mmproj-Q8_0.gguf	Qwen3-ASR 1.7B: its audio encoder with its projector	None: an unchanged copy of mradermacher's Q8_0	Apache 2.0	780b5b48dfbc23c219cc9d3da9fdaccc7da567fa661b3724d19bda50fbcfc8c3

Licences and notices

• Apache 2.0 (LICENSE-Apache-2.0.txt): EmbeddingGemma 2 and Qwen3-ASR. The files marked above as quantized or converted were modified from the originals in that way only.
• MIT (LICENSE-MIT-OmniShotCut.txt): OmniShotCut, Copyright (c) 2026 UVA Computer Vision Lab.
• "Gemma" and "EmbeddingGemma" are trademarks of Google LLC. EmbeddingGemma 2's model card asks that its use follow Google's Gemma Prohibited Use Policy.
• The models are provided as their authors provide them, without warranties of any kind.