Pydantic AI official releases
Pydantic AI 2.36.0 adds durable operations and a public backend API for durable execution engines.
Pydantic AI adds durable operations and a public backend API for third-party durable execution engines.
The complete AI news desk
The AI world, in one verified news room.
Models, agents, robotics, research, business and policy, with the original source attached. Search and move beyond the latest thirty updates.
Open the 30-day AI Change IndexLatest and archive
Open the fast flashes streamResearch
Unstructured official releases
Transform sign-up link added in version 0.27.5
The latest release includes a fix for hyphen handling in text partitioning.
Pydantic AI official releases
v2.35.3: Safe Download Cookies Limited to Original Hostnames
Version 2.35.3 released August 27, 2026. 2 bug fixes by @pydanty.
Microsoft DeepSpeed official releases
DeepSpeed reduces NCCL backend timeout from 30 to 10 minutes.
The v0.19.6 patch includes fixes for checkpoint rank selection and staticmethod handling on Python 3.9.
Dagster official releases
Dagster+ adds single-action wipe for dynamic partitions
Dagster+ now allows wiping and deleting dynamic partitions in one step, excluding multi-partitioned assets.
OpenAI Java SDK releases
The 4.54.0 version of the software makes function call output call IDs optional.
The 2026-08-27 release makes function call output call IDs optional.
OpenAI Go SDK releases
The 2026-08-27 release of v3.54.0 makes function call output call IDs optional.
The 3.54.0 release includes an optional call ID feature for API function outputs. The update also strengthens X.509 workload identity security.
OpenAI Ruby SDK official releases
Sorbet structured-output models added in v0.82.0
The 0.82.0 release includes optional Sorbet structured-output models. Function call output call IDs are now optional.
Hugging Face Hub Python SDK official releases
Hugging Face Xet downloads now reuse tokens, reducing API calls per file.
Xet downloads now reuse tokens, reducing API calls by up to 1,500 per minute.
Weaviate official releases
DigitalOcean module added to generative features in v1.38.13
The MCP stateless endpoint now refuses GET requests with a 405 status.
Swiss Data Science Center Renku official releases
Renku 2.20.0 adds project storage shared among all users of a project.
Renku 2.20.0 adds project storage shared among users, with administrators setting size limits.
Weaviate official releases
New TwelveLabs Marengo multimodal vectorizer added in v1.37.15
New modules added: multi2vec-twelvelabs and generative-digitalocean.
Pydantic AI official releases
Bedrock guardrail trace added to ModelResponse.provider_details
Version 2.35.1 released on August 26, 2026, includes bug fixes for provider details and Temporal responses.
Google Gemini CLI official releases
Nightly Release Fixes SSRF Vulnerability in MCP OAuth
A security fix for SSRF in MCP OAuth metadata discovery and authentication was implemented by @josebalius.
OpenAI Node SDK official releases
The 2026-08-26 release makes function call output call IDs optional.
The 2026-08-26 release makes API function call output call IDs optional. Streaming now respects cancellation during SSE response decoding.
OpenAI Python library releases
The 3.5.0 API now allows optional function call output call IDs.
The 3.5.0 release, dated August 27, 2026, includes an optional feature for function call output call IDs.
bitsandbytes official releases
Triton 4-bit quantize kernels updated in version 0.50.2.
ROCm 10.0 support added for Linux and Windows.
bitsandbytes official releases
Bitsandbytes Foundation updates main branch wheels to 1.33.7-preview
The latest pre-release wheel for bitsandbytes is version 1.33.7-preview. The installation command varies by platform.
Docling official releases
Docling 2.123.0 makes threaded docling-parse the default across its interfaces.
Threaded docling-parse becomes the default across the Docling SDK, CLI, service and extraction interfaces.
vLLM official releases
Kimi-K3 now supports ROCm with V2 model runner.
v0.28.0 includes 584 commits from 270 contributors, 76 of whom are new. The default max_num_batched_tokens is now 16384.
OpenCompass official releases
OpenCompass 0.5.4 adds LiteLLM as an AI gateway backend.
OpenCompass v0.5.4 adds LiteLLM as an AI gateway backend and includes new benchmarks for finance and code reasoning.
Pydantic AI official releases
LangChain migration skill added in v2.34.0
The 2026-08-24 release includes a new LangChain migration skill and GLM-5.3 support for ZaiModel.
Pydantic AI official releases
Temporal metric export frequency reduced in v2.35.0
The new release deprecates RunContext.capability_loaded and available_capability_ids. It also reduces the default Temporal metric export frequency.
Google Gen AI Python SDK releases
Software update 2.20.0 adds JPEG2000 support.
New audio transcription configuration mode added in v2.20.0.