Liquid AI published LFM2.5 Q4_0 checkpoints, produced through quantization-aware distillation (QAD). The method integrates quantization during training to preserve model quality at 4-bit precision. The checkpoints are available on Hugging Face, offering efficient deployment options for edge and resource-constrained environments.

LangChain released langgraph-sdk 0.4.3, adding a decrypt replacement result feature and supporting clearing cron end_time via update(end_time=None). The update includes dependency bumps and internal fixes across the LangGraph ecosystem.

Mojo, the Python-inspired language for GPU programming, is now open source. The compiler and toolchain were released under an Apache 2 license, following the 1.0 release last week. Originally planned as a Python superset, Mojo is now its own language, prioritizing GPU programming ease.

Mistral released CLI v0.3.0 with OIDC login/logout/whoami commands and CLI-based app evals. Mistral Vibe v2.24.2 adds a session picker, /log-level, /retry for ACP clients, worktree cleanup, faster startup, and multiple bug fixes.

OpenAI released Codex v0.148.0, adding TUI conversation export to Markdown, session forking with `codex exec fork`, draft prompts during startup, and Amazon Bedrock Runtime as a built-in provider. Bug fixes include model switching, session restoration, and sandbox restrictions failing closed on denied paths.

Hugging Face's Sentence Transformers library now supports multi-vector (late interaction) embedding models, enabling richer document representation. This update allows for more nuanced semantic search and retrieval, improving performance on complex queries.

openai-ruby v0.80.0 released on 2026-08-17 adds Amazon Bedrock Runtime support, plus API updates for Ultrafast tier, structured MCP, websocket errors, and separate websocket events. Bug fixes and chore improvements are also included.

OpenAI Ruby v0.79.0 released with HTTP response observability, Tapioca typing for structured outputs, WebSocket stream IDs, and Azure OpenAI v1 support. It also adds default headers, exposes request IDs, deprecates Sora video APIs, and includes bug fixes for model identifiers and timeout handling.

OpenAI released v1.1.0 of its Terraform provider, adding organization and project spend limits. The update also includes dependency bumps, performance caching for project user reads, and CI improvements. This release follows v1.0.0 and is available on GitHub.

OpenAI Ruby library v0.78.0 released, adding support for custom HTTP transports and hardening them. Bug fixes and dependency updates included. Released August 7, 2026.

OpenAI released Tart 2.35.0, a virtual machine tool for macOS. The update builds Tart on macOS 26 with Xcode 27, ensuring compatibility with the latest Apple operating system and development tools. This release includes a single pull request by torarnv.

Hugging Face released speech-to-speech v0.2.11, adding a chat-completions LLM backend (OpenAI /v1/chat/completions), support for text-only and out-of-band responses, a realtime web demo, and fixes for conversation races and Windows installation. Default Qwen3 TTS switched to GGML.

Microsoft Research released Orchard, an open-source framework for training and evaluating AI agents across task types. It reduces complexity and supports strong performance from smaller models by enabling researchers to reuse infrastructure, aiming to scale agentic AI research.

Hugging Face released a blog post titled 'State of Open Models: Summer 2026 Observations,' offering its assessment of the open-model ecosystem as of summer 2026. The post is published on the Hugging Face blog under this title, but no further details, data, or specific findings were provided in the available source content.

Hugging Face published a blog detailing how developers can use Strands Agents, LeRobot, and Hugging Face Storage Buckets together to record, train, and deploy robotics models from a single streaming data pipeline. The integration aims to simplify moving robot data from capture through training to deployment without switching platforms, using Hugging Face's storage infrastructure as the connective layer.

Allen Institute for AI introduced OlmoEarth embeddings, a new feature in OlmoEarth Studio that lets users export custom embeddings derived from satellite and geospatial data. These embeddings can be used for downstream analysis tasks, enabling researchers and developers to build custom models on top of pre-processed Earth observation data without starting from raw imagery.

NVIDIA is celebrating partners and open source communities advancing local AI throughout August, highlighting its latest open models, software, and tools. The initiative focuses on enabling developers to build, customize, and run increasingly capable AI agents locally, showcasing the growing ecosystem of models and applications.

Anthropic released Claude Code v2.1.233, adding GitLab merge request URL support, an opt-in forward_user_identity setting for apps gateway, memory cgroup limits for Bash commands, and a WebFetch cache TTL config. It fixes MCP v2 stream reconnection, notification hooks, and a Windows NTLM credential-leak vector.

Mistral AI released Vibe v2.24.0, introducing an admin config layer for shared/enforced settings, a 'Default' unpinned model option, server-side default model routing, and improved subagent handoff. It also fixes drag-and-drop in terminals and speeds up /resume listing.

Hugging Face released speech-to-speech v0.2.12, the final 0.2.x update. It introduces Smart Turn v3.2 for better endpointing, WebRTC transport for the OpenAI Realtime API, direct audio input for audio-capable LLMs, and an optional remote-LLM proxy. The browser demo also improved with device selection and user-turn feedback.

Hugging Face released harbor-hf 0.1.0, an initial PyPI release of a CLI tool for planning and executing campaigns on Hugging Face Jobs, Inference Endpoints, and Sandboxes. It includes durable evidence and endpoint cleanup checks. Install via `uv tool install harbor-hf`.

Hugging Face released Transformers v5.15.0, adding support for Meta's Muse Glimmer, a 30B multimodal model for agentic use cases, plus GraniteSWA, A.X-K1/K2, and Cosmos3 Edge. Breaking changes make kernels opt-in for linear attention models and require negative values for cache cropping.

LangChain released langgraph-checkpoint-postgres 3.1.2, fixing a bug in delta history walking for plain-value seeds and adding conformance suite tests. The release also bumps checkpoint to 4.2.0 and includes dependency updates.

LangGraph released checkpoint 4.2.0, adding an opt-in omit_expired feature to skip expired rows on read, fixing write collection in delta channel history, and updating dependencies including langsmith. The release includes multiple dependency bumps and code quality improvements.