Changelog

This page lists changes to the Vi SDK by version. Entries follow the Added / Changed / Fixed / Removed format.

---

v0.1.0b7

Security

• Upgraded the locked dependency set so every package flagged by an open security advisory sits at or above its patched version. Notable bumps: aiohttp 3.14.3, gitpython 3.1.59, pillow 12.3.0, starlette 1.6.0, torch 2.13.0, tornado 6.5.8, setuptools 84.0.0, h2 4.4.1
• Raised the pillow floor to >=12.3.0, since it is a shipped runtime dependency

Added

• Gemma 4 inference support, including native thinking and video inference
• Qwen 3.5 and Qwen 3.6 inference support
• New API resources: client.plans, client.folders, client.ontologies, client.deployments, client.deployment_projects, client.deployment_endpoints, client.asset_sources, client.intelliscribe, client.assets_with_objects, and client.annotation_import_sessions
• vi.__version__ is now exported from the top-level package
• Resumable model extraction

Fixed

• The response schema is now enforced when task_type is overridden
• Real-world HuggingFace checkpoints load correctly
• response_format no longer leaks, and the local deployment server returns real generated text
• Native thinking repaired for Qwen 3.6

---

v0.1.0b6

Added

• CRUD support for the flows and runs resources

Fixed

• Multi-GPU inference with qLoRA
• Video decoding, and more reliable asset progress tracking

---

v0.1.0b5

Added

• NVIDIA Cosmos Reason2 support, for both local inference and NIM deployment
• Video inference support
• Localhost API deployment: the OpenAI-compatible local inference server
• HuggingFace dataset import integration
• Video and multipart upload support for assets

Changed

• Migrated inference to Transformers v5

Fixed

• Interrupt handling in wait_until_done
• Error messages are now propagated for long-running operations

---

v0.1.0b4

Fixed

• Fall back to HTTP/1 when using multiple concurrent connections, which fixes asset downloading

---

v0.1.0b3

Fixed

• Inference execution inside Jupyter notebooks

---

v0.1.0b2

Added

• DeepSeek-OCR support
• LLaVA-NeXT support

Fixed

• Inference validation

---

v0.1.0b1

Initial release

Added

• vi.Client: entry point for all SDK operations, with secret_key and organization_id authentication
• Dataset management: client.datasets.list(), client.datasets.get(), client.datasets.create()
• Asset operations: client.assets.upload(), client.assets.list() with concurrent upload support and progress tracking
• Annotation workflows: client.annotations.list(), client.annotations.create() for phrase grounding and visual question answering (VQA) tasks
• Model training: client.runs.list(), client.runs.get() for monitoring training runs
• Local inference via vi.inference.ViModel (supports Qwen2.5-VL, InternVL 3.5, Cosmos Reason1, and NVILA models)
• Dataset loaders: vi.dataset.loaders.ViDataset for loading downloaded datasets and iterating training/validation splits
• NIM deployment: vi.deployment.nim.NIMDeployer, NIMPredictor, and NIMConfig for GPU-accelerated container inference
• Typed exceptions: ViError, ViAuthenticationError, ViNotFoundError, ViValidationError
• Built-in pagination via .all_items() and page-by-page iteration
• client.help() and per-resource .help() methods for in-session reference
• Optional install targets: vi-sdk[inference], vi-sdk[jupyter], vi-sdk[deployment], vi-sdk[all]
• Python 3.10–3.14 support on Linux, macOS, and Windows

---

Related resources