VextLabsinc/juwel-emerald
VextLabsinc/juwel-emerald is a 33.4 billion parameter multimodal language model based on the Qwen3-VL architecture family, developed by Vext Labs. It features a 32,768-token context length and a 27-layer vision tower, enabling it to process both text and image inputs. This model is designed for conditional generation tasks involving image and text, leveraging its large parameter count and multimodal capabilities.
Loading preview...
Overview
VextLabsinc/juwel-emerald is a 33.4 billion parameter multimodal language model from Vext Labs, built on the Qwen3-VL architecture. It is designed for conditional generation, integrating both text and image processing capabilities. The model features a substantial 32,768-token configured context length and a 27-layer vision tower, allowing it to handle complex multimodal inputs. This release consists of 14 safetensors shards, totaling 66.7 billion tensor bytes, and is mirrored from Vext Labs' public Cloudflare R2 release.
Key Capabilities
- Multimodal Processing: Utilizes a Qwen3-VL architecture with a dedicated 27-layer vision tower for image understanding.
- Large Context Window: Supports a 32,768-token context length for processing extensive inputs.
- Conditional Generation: Capable of generating text based on combined image and text prompts.
Usage Considerations
- Hardware Requirements: Due to its large size (33.4B parameters, BF16 artifacts), appropriate hardware and sharding are necessary for efficient loading and inference.
- Verification: The release ensures structural consistency and byte integrity, but does not guarantee output quality, training data provenance, safety, or production fitness. Users should review the configuration and receipt before deployment.
- License: Distributed under the Apache License 2.0, which governs the mirrored files. "Qwen" refers to architectural lineage only and does not imply endorsement.