Cortex-Mini-1-Preview

Cortex-Mini-1-Preview is a post-trained derivative of Qwen/Qwen3.8-27B, released by CortexLM.

This repository is a fork of the official Qwen3.8-27B checkpoint. We start from that open-weight native vision-language model and continue with Cortex post-training (Relearn). The goal is to improve the model on held-out tasks without collapsing general capability or overfitting public eval splits.

Preview. Weights, training recipe, and numbers in this card will move. Treat this page as the public home for the Cortex Mini 1 line, not a frozen release.

Thanks

This work exists because the Qwen team open-sourced Qwen3.8-27B under Apache 2.0.

Thank you to the Qwen Team and Alibaba Cloud for the architecture, the training, the weights, and the docs. Cortex-Mini-1-Preview inherits their model; our contribution is post-training on top. If you use this checkpoint, please also cite Qwen3.8 (see Citation).

What this is

Qwen3.8-27B already accepts image and video input. Cortex-Mini-1-Preview does not bolt on a separate vision encoder. Miners and researchers post-train this 27B checkpoint itself.

License

The base model is licensed under Apache License 2.0, Copyright 2026 Alibaba Cloud. That license allows commercial use, modification, and redistribution, including this fork.

This repository redistributes Qwen3.8-27B weights and configuration as a derivative work:

• The Apache 2.0 license text ships with the repo (LICENSE).
• Attribution to the Qwen Team / Alibaba Cloud is retained (this card + NOTICE).
• Files we change are marked as such. The model card, banner, and Cortex post-training artifacts are new.

You must keep the license, attribution, and NOTICE if you redistribute further. "Qwen" remains a trademark of the original authors; this project does not claim that mark.

Full terms: Apache License 2.0 and the LICENSE file in this repo.

Model overview (inherited from Qwen3.8-27B)

These architectural facts come from the Qwen3.8-27B model card. Cortex post-training does not change the network shape.

• Type: Causal language model with vision encoder (native VLM)
• Training stage (base): Pre-training and post-training by Qwen
• Training stage (this repo): Additional Cortex post-training on the Qwen checkpoint
• Parameters: 27B
• Hidden size: 5120
• Layers: 64
• Context: 262,144 tokens natively, extensible to 1,000,000 with YaRN
• Modalities: Text, image, video

For serving details, sampling defaults, thinking mode, and YaRN flags, follow the Qwen card. Those recipes still apply.

Intended use

• Research and development on Cortex Relearn (post-train a capable open VLM without wrecking general skills).
• Downstream chat, coding, agent, and vision-language workloads that already run on Qwen3.8-27B.
• Further fine-tuning under Apache 2.0.

Out of scope

• Do not present this preview as a drop-in Qwen Cloud / Qwen3.8-Max replacement.
• Do not treat unpublished scores as Qwen's official numbers.
• Do not strip license or attribution.

How to use

Load the repo id once weights are in place:

from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "CortexLM/Cortex-Mini-1-Preview"
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)

For production serving, use vLLM, SGLang, or TokenSpeed with the local directory of the downloaded weights, not a guessed path. See Qwen's vLLM recipe and SGLang cookbook.

Thinking mode is on by default in the Qwen3.8 stack (enable_thinking, preserve_thinking, reasoning_effort). Keep those defaults unless you know you want instruct/non-thinking sampling.

Evaluation

Preview: we do not publish a public leaderboard on this card yet. Cortex scores live training on a private holdout. Public splits are informational. General-capability checks run off the visible reward so they are not an overfitting target.

When a numbered release ships, this section will get frozen metrics and the eval recipe.

Limitations

• Preview checkpoint: APIs, tokenizer extras, and chat template can still change.
• Post-training can move behavior away from the Qwen defaults (style, refusal, tool use). Compare against Qwen/Qwen3.8-27B before you swap it in.
• Long-context and hour-scale video settings are inherited from Qwen; we have not re-tuned those knobs here.
• As with any open VLM, outputs can be wrong, biased, or unsafe. You are responsible for downstream filters.

Files of ours vs Qwen's

Citation

If you use Cortex-Mini-1-Preview, please cite both this work and Qwen3.8:

@misc{cortex-mini-1-preview,
  title  = {Cortex-Mini-1-Preview},
  author = {CortexLM},
  year   = {2026},
  url    = {https://huggingface.co/CortexLM/Cortex-Mini-1-Preview}
}

@misc{qwen38,
  title  = {{Qwen3.8-Max}: A New Bar for Coding and Cowork},
  url    = {https://qwen.ai/blog?id=qwen3.8},
  author = {{Qwen Team}},
  month  = {August},
  year   = {2026}
}

Links

• Base model: Qwen/Qwen3.8-27B
• Qwen collection: Qwen3.8 on Hugging Face
• CortexLM: huggingface.co/CortexLM
• Cortex: github.com/CortexLM/cortex