By the way, all the features listed below you can leverage without worrying about scalability or rate limits. Our infrastructure automatically scales with your business growth, ensuring seamless performance from prototype to production.

Vision models: bringing sight to your AIVision models: bringing sight to your AI

This update introduces vision-language models that enable your applications to understand and interact with visual content. From image captioning to product recognition, these models let you add a new dimension to your AI capabilities — perfect for holiday product catalogs or seasonal marketing campaigns.

New vision models include:

• Qwen2-VL-72B-Instruct: High-performance for complex visual tasks.

• Qwen2-VL-7B-Instruct: Lightweight yet versatile.

• LLaVA-v1.5-13b and LLaVA-v1.5-7b: Specialized in image captioning and visual Q&A.

Expanded language model portfolioExpanded language model portfolio

We’ve broadened our language model offerings to meet diverse use cases — from complex reasoning to multilingual scenarios. Whether you’re refining product recommendations before the holiday rush or planning next year’s multilingual marketing campaigns, there’s a model that fits your needs.

New language models include:

• Meta Llama-3.3-70B-Instruct: Access Meta’s latest state-of-the-art model featuring a 131K context window and support for 8 languages including English, German, French, Italian, Spanish, and Thai. Perfect for enterprise applications requiring top-tier performance.

• dolphin-2.9.2-mixtral-8×22b: Built for conversational AI and coding tasks, featuring a 64K context window. Excels at following detailed instructions and handling complex dialogues. Perfect for chatbots and coding assistants.

• Phi-3.5-MoE-instruct: Advanced mixture-of-experts model balancing high performance with efficient resource usage. Ideal for production applications requiring reliable performance.

• Phi-3.5-mini-instruct: Compact version delivering quick responses for applications where speed is essential.

• Qwen2.5-1.5B-Instruct: Fast, lightweight model perfect for testing and rapid prototyping.

• Llama 3 Series (Med42-8B, 3.2 1B, 3.2 3B): Range of models for different needs, from specialized medical applications to general NLP tasks.

All models are available through our unified API, with both base and fast inference options to match your performance needs.

New embedding models for RAGNew embedding models for RAG

We’re adding three embedding models to strengthen your retrieval-augmented generation pipelines. Perfect for building knowledge bases, advanced semantic search engines or contextual chatbots — just in time to organize year-end data reviews.

New embeddings:

• BGE-ICL: A leading open-source embedding model.

• e5-mistral-7b-instruct: Instruction-tuned for richer context.

• bge-multilingual-gemma2: Multilingual support for global applications.

LoRA hosting: a simpler, usage-based approachLoRA hosting: a simpler, usage-based approach

We’re taking a different approach compared to other providers: pure usage-based pricing, no fixed costs and zero overhead — with no dedicated instances, monthly commitments or infrastructure management.

Here’s how it works:

• Bring your LoRA weights: Simply upload your pre-trained LoRA model to Nebius AI Studio.

• Start running inference: No environment setup or scaling decisions needed.

• Pay only for what you use: Our per-token pricing ensures you don’t pay when your models sit idle.

This is ideal for teams that have already fine-tuned their models elsewhere and now need a hassle-free, cost-effective way to deploy and run them — especially during high-demand periods, without paying for unused capacity.

How to get started with LoRA:

1. Go to the LoRA section in the AI Studio UI.

2. Click “Request Hosting.”

3. Submit your model details.

4. Our team will handle the rest.