Vision-language fine-tuning

Instruction fine-tuning for vision-language models (VLMs) that reason jointly over images and text. Users can now fine-tune multimodal models using datasets combining visual inputs and natural language through the same fine-tuning workflows used for text-only models.

Capabilities:

• Upload and validate vision-language datasets with messages-based schema and image support
• Fine-tune supported VLMs: Qwen2.5-VL-7B-Instruct and Llama-3.2-11B-Vision-Instruct
• Automatic schema validation and seamless job creation
• Standard instruction tuning workflow with no separate multimodal configuration required

ℹ️

This release supports instruction tuning only. Reinforcement learning and preference tuning are not yet supported for vision-language models.

Tools library

Centralized workspace for managing agent tools across your organization. The tool library serves as a single source of truth where users can create, update, delete, and duplicate tools independently of agent configuration.

Capabilities:

• Create and manage web search and file search tools with custom configurations.
• View code interpreter and custom function tools (read-only).
• Duplicate existing tools to create variations with modified configurations.
• Automatic propagation of tool updates to all agents using that tool.
• Select pre-created tools during agent creation or create new tools inline.
• Query tools by type and status, and view which agents are using specific tools.

Tool changes automatically redeploy linked active agents to reflect updates across your organization.

Ingestion insights

Real-time visibility into file ingestion status across Alignment and VectorDB endpoints. Every uploaded file receives its own persistent ingestion record that tracks progress through conversion, alignment, and vectorization.

Capabilities:

• Per-file status tracking (queued, running, completed, failed) with queue position and timestamps.
• Plain-language error messages with suggested fixes.
• Structured metadata through file_records view showing which files are processing, completed, or blocked.
• Consistent response schema across Alignment and VectorDB endpoints.

This visibility layer enables faster self-service debugging and reduces support escalations during onboarding, proofs of concept, and production workflows.

Multi-node fine-tuning

Distributed training across multiple compute nodes for improved fine-tuning performance and scalability. Multi-node fine-tuning distributes training jobs across multiple physical nodes (each with 8 GPUs), enabling parallel execution rather than single-machine training.

Capabilities:

• Support for 1–4 nodes (up to 32 GPUs total)
• Configure node count via n_node parameter in InfrastructureConfig
• Automatic distributed orchestration and synchronization
• Immediate validation of supported n_node / n_accel combinations

This enables reduced training time for larger datasets, better compute allocation matching dataset size and training needs, and establishes the foundation for future support of larger models.