PyTorch 2.6 changed torch.load to behave as weights_only=True when callers do not provide a custom pickle_module. Plain tensor state dictionaries should continue loading, while serialized modules, tensor subclasses, NumPy objects, and other custom globals can now fail until the caller chooses an explicit trusted-load path or allowlists required types.
OpenAI has shut down the Assistants API and directs integrations to the Responses API, with Conversations used for persistent state. The migration changes the core object model from Assistants, Threads, Runs, and Run Steps to Prompts, Conversations, Responses, and Items.
NumPy 2.0 introduced the project’s first major ABI break since the 1.x line, so extension binaries built against NumPy 1.x may fail to import under 2.0. It also removed old Python and C APIs and adopted new type-promotion behavior, creating migration work even for some projects without compiled code.
MLflow 3 promotes logged models to first-class entities, changes model logging from artifact_path to name, and stores model artifacts outside the run-artifact hierarchy. It also removes Recipes, several model flavors and parameters, the deployment-server command, and portions of the older evaluation API.
Transformers v5 made PyTorch the sole model backend, removed long-deprecated surfaces, and replaced the slow-versus-fast tokenizer split with a new backend architecture. The major release also added dynamic weight conversion and moved to huggingface_hub>=1.0, bringing HTTPX exceptions, environment-based proxy configuration, and cache-variable changes into the migration.
Hugging Face Transformers 5.0 / v5 migration · Source checked
Google replaced the Interactions API outputs array with typed steps and consolidated output controls under a polymorphic response_format. Its current migration timeline marks the legacy schema removed on June 8, 2026, after the new schema became the default on May 26.
Google shut down the Gemini 2.0 Flash and Flash-Lite model IDs, including their -001 variants, on June 1, 2026. The current lifecycle table recommends Gemini 3.6 Flash for the Flash IDs and Gemini 3.1 Flash-Lite for the Lite IDs.
For models released after Claude Opus 4.6, Anthropic accepts temperature only at 1.0 and top_p only at 0.99 or higher, while any explicit top_k is rejected with HTTP 400. The Python SDK version 1.0 and later rejects these fields locally with TypeError.
Claude Messages API; models released after Opus 4.6 · Source checked
No changes match these filters. Try another provider or clear the watchlist filter.