CortexLM/Cortex-Mini-1-Preview

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 31, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Cortex-Mini-1-Preview is a 27 billion parameter native vision-language causal model developed by CortexLM, derived from Qwen/Qwen3.8-27B. This model undergoes Cortex post-training (Relearn) to enhance performance on held-out tasks without compromising general capabilities or overfitting public evaluation benchmarks. It supports text, image, and video inputs, inheriting its architecture and multimodal capabilities directly from the Qwen3.8-27B base model. The primary goal of this continued training is to improve the model's effectiveness while maintaining its robust foundation.

Loading preview...

Cortex-Mini-1-Preview: Post-Trained Vision-Language Model

Cortex-Mini-1-Preview is a 27 billion parameter native vision-language model developed by CortexLM. It is a post-trained derivative of the Qwen/Qwen3.8-27B model, inheriting its core architecture and multimodal capabilities, including support for text, image, and video inputs. The model's continued training, termed "Relearn" by CortexLM, focuses on improving performance on specific held-out tasks.

Key Capabilities & Features

  • Multimodal Input: Natively processes text, image, and video data without requiring separate vision encoders.
  • Continued Training: Undergoes Cortex post-training to enhance performance on specific tasks while preserving general capabilities.
  • Base Model Inheritance: Retains the architectural facts and robust foundation of the Qwen3.8-27B model, including its 27B parameters, 64 layers, and a native context length of 262,144 tokens (extensible to 1,000,000 with YaRN).
  • Apache 2.0 License: Licensed under Apache License 2.0, allowing for commercial use, modification, and redistribution.

Intended Use Cases

  • Research & Development: Ideal for exploring Cortex Relearn methodologies, focusing on post-training capable open VLMs without degrading general skills.
  • Multimodal Workloads: Suitable for downstream chat, coding, agent, and vision-language applications that currently utilize Qwen3.8-27B.
  • Further Fine-tuning: Provides a strong base for additional fine-tuning under the permissive Apache 2.0 license.

Finetunes
1 models