A unified training engine for multimodal models, supporting vision-language, diffusion, and MoE architectures with FSDP2, Muon, and Liger kernels.
EvolvingLMMs-Lab/NEO
Open-source native vision-language models (NEO series) that process pixels and words in one encoder-free dense architecture, with 2B and 9B releases.
- Estrellas
- 903
- Forks
- 0
- Issues abiertos
- 0
- Último commit
- hace 3 meses
Funciones clave
- 01Encoder-free dense architecture that unifies pixel and word encoding in one model.
- 022B and 9B model weights released across training stages.
- 03Evaluation guide and code built on VLMEvalKit.
- 04Training code for NEO models in a dedicated guide.
- 05Covers single-image, multi-image, video and spatial understanding tasks.
Casos de uso
- Researching native or encoder-free multimodal model design.
- Benchmarking vision-language models on image, video and spatial tasks.
- Fine-tuning a compact 2B or 9B VLM on custom image-text data.
Temas
Preguntas frecuentes
- ¿Qué es NEO?
- Open-source native vision-language models (NEO series) that process pixels and words in one encoder-free dense architecture, with 2B and 9B releases.
- ¿NEO es gratuito? ¿Qué licencia tiene?
- NEO es open source bajo la licencia Apache-2.0. Revisa los términos de la licencia antes de usarlo en productos comerciales.
- ¿NEO sigue mantenido?
- Actualizado en los últimos 6 meses. El último commit de NEO fue el 27 jul 2026.
- ¿Para qué puedo usar NEO?
- Researching native or encoder-free multimodal model design. Benchmarking vision-language models on image, video and spatial tasks. Fine-tuning a compact 2B or 9B VLM on custom image-text data.
- ¿Dónde consigo NEO?
- El código fuente y las instrucciones de instalación están en GitHub: https://github.com/EvolvingLMMs-Lab/NEO
Resumen generado por IA a partir del README del proyecto. Las estadísticas de GitHub se actualizan a diario.
Proyectos relacionados
A unified multimodal model that handles many computer vision tasks by generating text, images, or both from natural-language instructions.
Visual RAG toolkit that renders web pages and PDFs as screenshots and retrieves over the images, keeping tables and charts intact.
Open-source native multimodal models that handle image understanding, text-to-image generation, and image editing within one unified architecture.
A 9B-parameter open-weight multimodal model for visual understanding, spatial reasoning, tool use, and embodied-AI research, built on Qwen3.5 and C-RADIOv4-H.
Converts PDFs, Office files, scans, screenshots and diagrams into Markdown by sending pages to your own OpenAI-compatible vision model.