AI.info
AI Tools Updates — Page 108
What changed in the AI tools this week — dated, from the makers’ own changelogs.
En anglais
- Hugging Face — Holo4: powering generalist computer-use agents
Holo4 powers generalist computer-use agents.
- ElevenLabs — Eleven v4 and Eleven v4 Turbo
Eleven v4 provides higher-quality, more expressive speech and improved voice cloning across more than 90 languages; Eleven v4 Turbo is for real-time applications.
- LiteLLM — v1.103.0
Config file ownership, Fuse and Capability routing, gateway hardening
- Unstructured — 0.27.10
HTML partitioning now extracts definition lists instead of discarding them.
- LiteLLM — Config file ownership, Fuse and Capability routing, gateway hardening
- InvokeAI — Version 6.14.2
Fixes bugs in version 6.14.1, adds custom font import for Canvas text, and transfers an image’s style to krea-2 using a reference image.
- Letta — Release v0.33.4
Updated remote terminal evidence checks and stopped child sends from retaining parent runtimes.
- LiteLLM — v1.103.0: Config File Ownership, Fuse Routing & Gateway Hardening
The config file now owns every setting it declares; Auto Router adds a capability classifier, Fuse v2 forecasting, per-model Fast mode.
- Speakeasy SDK Generation — Directory groups grant roles on their own, and Slack accounts resolve to the people behind them
Admins can map directory groups and attributes to roles, and sync Slack workspaces and map members to organization identities.
- Speakeasy SDK Generation — Remote MCP servers choose who they act as, and a workload fleet can be admitted by one wildcard
Remote MCP servers can select an identity mode, and permitted wildcard admissions can cover workload subjects beneath a path.
- Speakeasy SDK Generation — The latest models reach the pickers, and published plugins work on OpenCode 2 again
Current OpenRouter models are selectable, plugins support OpenCode 2, and deployments can use additional first-party hosts.
- Speakeasy SDK Generation — Risk policies can target chosen MCP servers, gateways, and tools
Risk policies can target selected MCP servers and tools, and admins can disable Shadow AI scans in fleet configuration.
- InvokeAI — Version 6.14.2 Latest
You can now import custom fonts for use with the Canvas text tool and transfer an image’s style to krea-2 by providing a reference image.
- Gram — The latest models reach the pickers, and published plugins work on OpenCode 2 again
New models are available in model pickers, plugins support OpenCode 2, and deployments can use extra first-party hosts.
- Gram — Risk policies can target chosen MCP servers, gateways, and tools
Risk policies can target selected MCP servers and tools, and admins can disable Shadow AI scans for a fleet.
- Gram — Remote MCP servers choose who they act as, and a workload fleet can be admitted by one wildcard
Remote MCP servers can choose an identity mode, and workload fleets can be admitted with an allowed trailing wildcard.
- Gram — Directory groups grant roles on their own, and Slack accounts resolve to the people behind them
Directory groups and attributes can grant roles, and Slack accounts can be mapped to people in the organization.
- Unsloth — Flash-Attention2, Causal-Conv1D, Mamba_SSM Binaries
Prebuilt Linux x86_64 CUDA 13 wheels are available for flash-attn, causal-conv1d, and mamba-ssm on PyTorch 2.13 and 2.14 with Python 3.13.
- Riverside.fm — Q&A AI Moderator
Let AI handle the busywork of live Q&A so you can focus on running the webinar.
- Riverside.fm — Riverside now works in Safari!
Mac users can log in, record, edit, and publish without switching to Chrome or Edge.
- Riverside.fm — Regenerate your transcript
Regenerate a recording's transcript from the Editor in a language you choose.
- Unstructured — 0.27.9
Chunking retains table rowspans across cell splits.
- Tines Agents — 🔐 View connector destination changes in audit logs with URL pattern and host metadata details
Audit logs show connector destination changes with URL pattern and host metadata details.
- Arize Phoenix — Run high-volume structured evals with Jev-as-a-Judge
You can evaluate traces with TypeSafe AI’s Jev decision model from Evaluator Hub.
- Arize AX — Run high-volume structured evals with Jev-as-a-Judge
Evaluate traces with TypeSafe AI’s Jev decision model from Evaluator Hub and attach it to an online eval task.
- Factory Droids — CLI v0.228.0, Desktop v0.185.0: Command palette rebuild, migrate command, mid-message skills, in-place compression, plus terminal and app fixes
Added migration file viewing, delegation audit events, image generation model choice, image expansion, skills anywhere, and /migrate.
- Mixedbread Rerank — Scope API Keys by Store Tag
API keys can be scoped to Stores with a specific tag, including Stores tagged later.
- Tines Workflow Automation — 🔐 View connector destination changes in audit logs with URL pattern and host metadata details
Audit logs now show connector destination changes with URL pattern and host metadata details.
- Fireworks AI — Serverless deprecation: DeepSeek V4 Pro (0813), DeepSeek V4 Flash (0731), and related models
The models below are no longer available on public serverless. Dedicated deployments are unaffected.
- Mixedbread Embed — Scope API Keys by Store Tag
API keys can now be scoped to all Stores with a specific tag, including Stores tagged later.
- Keploy — v3.6.77
Keeps the agent token private, supports snap Docker, and stops false Kubernetes auth warnings.
- Greptile — Review Tiers
Choose how much work Greptile puts into each review: Base, Plus, Apex, or Auto.
- Relevance AI Workforce — Trigger status page and failure alarms
The Trigger status page lists every trigger in a project with its current status and latest runs, and can notify you when one fails.
- Open Interpreter — rust-v0.0.51
Fixed mode-selected macOS artifacts for OIX 0.0.51.
- UiPath Maestro — Autopilot as a coding agent for Maestro Orchestrate projects in Studio Web — Preview
Autopilot plans, builds, and edits a project from a plain-language description across Flow, BPMN, and Case projects.
- RunPod Serverless — Unified credentials page
API keys, S3 API keys, SSH public keys, container registry credentials, and secrets now live together on a single Credentials page in the console.
- Anima — Claude Docs: What is Claude Docs, Pros, Cons, and Alternatives
Claude Docs launched in beta alongside Claude Slides and the merger of regular chat and Cowork into one Claude experience.
- Abnormal Inbound Email Security — Weekly Detection Enhancements - Week of September 25
Added internal-account compromise detections; improved precision of subject-matching detection, reducing false positives
- Pulumi Neo — Change your components, keep your infrastructure
You can change the component managing existing infrastructure without replacing that infrastructure.
- Botpress — September 25, 2026
Admins can configure auto-close in hours; added HubSpot ticket actions; SLAs count human replies toward response targets.