Hugging Face ❤️ TorchTPU: TPUs, Natively
From the official blog (huggingface.co), published . View original
Hugging FaceOfficial blogAI news
Hugging FaceOfficial blogHow two hackathon teams built tools for scientists and small businesses with marimo
Their project connected agent-proposed experiments, physical laboratory runs, and analysis.
marimoOfficial blogImpeccable v4.7.0
Impeccable skill v4.7.0 can critique and audit Electron, NW.js, CEF or WebView2 apps started with a debugging port, scanning the open window without changing it. Two new typography checks arrive: straggler flags a paragraph ending on a single word, and unbalanced-heading flags a heading whose last line is a stub. The skill description loaded into every session is 30% shorter and now recognises requests to write a DESIGN.md. Decision and review pages follow the OS dark mode, and in monorepos only workspaces that render a UI are offered as targets.
ImpeccableGitHub ReleasesEasydict v2.25.0
Easydict 2.25.0 adds an Anki Connect integration with its own settings sheet. The DeepL client version has been updated to restore translation compatibility, and DeepL gains Hebrew and Thai. Gemini has moved to the OpenAI format. Query window resizing and text rendering are now more stable.
EasydictGitHub ReleasesClef is available on Ollama's cloud
Clef turns a state and a schema of typed questions into decisions.
OllamaOfficial blogDocumenso v2.21.0
Documenso v2.21.0 disables accounts when they are reported and shows a warning for new accounts. Multi-signer completion in v1 now works when optional fields are left empty.
DocumensoGitHub ReleasesAI Workspace Permissions: Roles, Data Access, and Checks
From the official blog (affine.pro), published . View original
AFFiNEOfficial blogOpen Source Notion Alternative: The Tradeoffs Most Lists Miss
From the official blog (affine.pro), published . View original
AFFiNEOfficial blogDigital Workspace Solutions 2026: The All-in-One Guide
From the official blog (affine.pro), published . View original
AFFiNEOfficial blog4 Collaborative Online Workspaces Compared for 2026 Teams
From the official blog (affine.pro), published . View original
AFFiNEOfficial blogshadcn/ui v0.3.0
Version 0.3.0 of @shadcn/helpers adds a { preliminary: true } option to the tool handle's output() for streaming preliminary tool output.
shadcn/uiGitHub ReleasesNTT Docomo Business trials conversational AI in Aizuwakamatsu with ElevenLabs
Central to the partnership is CONN, NTT Docomo Business's AI assistant, powered by ElevenAgents.
ElevenLabsOfficial blogElevenLabs’ commitment to security, privacy, and responsible AI: certified and renewed
But trust isn't a claim a company gets to make about itself, it has to be independently verified.
ElevenLabsOfficial blogCursor Cloud Agent Canvases and Environment Improvements
Cursor cloud agents can now build interactive canvases as part of their work and share a finished one with the team by link. The Environments API is now in beta for programmatically creating and managing the cloud environments agents run in. Up to 300 repositories now fit in one cloud agent environment, and Cursor says large environments load 5x faster.
CursorProduct release notesReflection's Beam is freely available on Ollama for a limited time
Beam is built for workflows that involve more than a single response: writing and maintaining code, using tools across multiple steps, and working through complex tasks over time.
OllamaOfficial blogPolar v3.0.0
Polar CLI v3.0.0 renames polar listen and polar trigger to polar webhook listen and polar webhook trigger, and adds polar auth token for creating an organization access token through the dashboard. Every command gains a --json flag, and API commands can cut their output down with --fields.
PolarGitHub ReleasesFolo v0.6.0
Folo Mobile v0.6.0 tells you when a newer version is on the App Store or Google Play, and two-factor setup gains a Copy setup key action. YouTube videos inside articles now appear as thumbnails that open the video rather than a player error. Scrolling no longer marks only the first entry as read. Setting an avatar and importing an OPML file, which always failed, now work.
FoloGitHub Releases 1 more updateT3 Code v0.1.0
T3 Code v0.1.0 is the first stable release of Orchestrator V2, which rebuilds how agent turns start, stop, queue and resume, along with subagent tracking and saved history. With delegate_task, an agent can launch child agents on any provider or model and wait for their results, any thread can be forked, and queued messages survive restarts. On first launch, V2 makes a one-way copy of the existing database and imports threads while the old version keeps its own copy. A V2 server cannot talk to a V1 app, the mobile app must move to 2.0, and each provider has a minimum version.
T3 CodeGitHub ReleasesImpeccable v4.6.0
Impeccable cli v4.6.0 installs on Node 18.19 or later instead of demanding Node 22.18, so npm no longer falls back to the old CLI. Its detect command runs engine 0.1.15 with the finding fixes from skill v4.6.0, and install can now target Oh My Pi.
ImpeccableGitHub Releases 1 more updateSSML tags explained: How to control pauses, pronunciation, and pacing in text to speech
From the official blog (elevenlabs.io), published . View original
ElevenLabsOfficial blogAudio deepfake detection: How it works & how to verify ElevenLabs audio
From the official blog (elevenlabs.io), published . View original
ElevenLabsOfficial blogVercel integration
The integration writes the credentials.
GraphifyOfficial blogGoogle DeepMind and Singapore Government collaborate to build fine-tuned AI model for scam detection
From the official blog (deepmind.google), published . View original
Google DeepMindOfficial blogAnythingLLM v1.18.0
AnythingLLM v1.18.0 is mostly a bug-fix release, and Kubernetes probes now point at the server's /api/ping endpoint. In half-hour time zones, the minute of an hourly schedule is converted to UTC correctly. Scheduled jobs stop running after a switch to multi-user mode. Confluence load failures are reported as errors instead of showing no pages.
AnythingLLMGitHub ReleasesHindsight v1.0.0
Hindsight v1.0.0 can register as OpenClaw's native memory provider, and it removes the OpenClaw agent-sdk package and the agent_knowledge_* tools as a breaking change. Every integration now identifies itself with its own User-Agent. A store that sheds a request answers 503 with Retry-After instead of 500. The release also fixes several coding-agent integration bugs, including lost chunks in the Claude Code integration and Cursor hooks on Windows.
HindsightGitHub ReleasesBaroness Campbell gives the UK Parliament’s first ever AI-voiced speech with ElevenLabs
“I used AI to speak my own words, not to put words into my mouth. The speech, the arguments and the views are entirely mine,” Baroness Campbell said.
ElevenLabsOfficial blogIntroducing Hindsight Memlake: serving 100M Memories at <100ms p99 on object storage
Changing the storage engine didn't cost retrieval quality.
HindsightOfficial blogRun AI agents safely with Ollama and Fly.io
The Sprite itself only knows a gateway URL and a connection ID.
OllamaOfficial blogPaperclip v2026.1012.0
Paperclip v2026.1012.0 lets Claude Code, Codex, OpenCode and other MCP clients connect as the signed-in user to review work and delegate tasks, behind an experimental setting. The Connections catalog can import accounts already linked in Composio and Arcade, and Codex runs continue after a server restart. Workspace restore locks switch to SQLite file locks, so every old process must be stopped before upgrading or rolling back. Stored tool grants are enforced exactly as written instead of falling back to broader policies, so custom grants need review after the upgrade.
PaperclipGitHub ReleasesPageIndex v0.4.0
PageIndex v0.4.0 makes the Flash engine the local default, building the tree from layout statistics without an LLM and using LLMs only for node summaries. PageIndexLocalClient matches the cloud client, with submit, tree, page content, chat and agent tools in both modes, and local mode needs no server, vector database or API key. The SDK wraps agent tools and configs for OpenAI, Anthropic and Claude, and chat() can stream or drive the OpenAI Responses and Anthropic Messages protocols directly. With no model set, local mode indexes with gpt-6-luna and answers with gpt-6-sol.
PageIndexGitHub ReleasesGraphify passed 125,000 GitHub stars
From the official blog (graphify.com), published . View original
GraphifyOfficial blogThe Last Shift: Qwen Image and LTX 2.5 Video with ComfyUI on a Mac
From the official blog (rapidmlx.com), published . View original
Rapid-MLXOfficial blogOpenCode v1.19.0
OpenCode v1.19.0 can upgrade to OpenCode 2 through opencode upgrade --major, and the TUI command /update installs updates or performs the upgrade, so updates are no longer automatic. The release allows the OpenAI ultrafast service tier and validates PTY connection origins and tickets before checking whether a PTY exists.
OpenCodeGitHub Releases