Downloadable conversational agent model supports tool calling, configurable reasoning, and local deployment using approximately 4GB in selected workloads.
New AI tools & model updates
New tools, new models, and major upgrades released in the last 14 days.
Coding model for Plus or Moderato subscribers and above; Pro or Allegretto tiers unlock million-token context. Thinking mode is required.
Faster coding model for Pro or Allegretto subscribers and above, with million-token context, mandatory thinking, and approximately twice the regular model's quota usage.
Consolidated Teams app connects notifications with Copilot investigation and pull requests; public preview requires a paid Copilot plan and enabled cloud sandboxes.
AI infrastructure monitoring platform correlates GPU, network, storage, and workload telemetry, with conversational investigation and recommended remediation.
Generally available training session APIs enable custom post-training loops without managing GPUs, initially in Northern Virginia, Oregon, and Mumbai.
Adds Bedrock inference API compatibility and configurable fallback routing between models or providers across supported AWS regions outside China.
Version 2.1.309 fixes sandbox boundary enforcement, MCP configuration protection, Bash permission checks, and worktree isolation guards.
Version 1.0 provides stable APIs, session-aware routing, traffic prioritization, and coordinated upgrades for distributed model inference.
NVIDIA announces general availability of Isaac Lab 3.0 for multi-GPU robot policy training using Newton and PhysX physics engines.
Reference architecture combines agent runtime isolation, monitoring, and hardware policy enforcement across enterprise infrastructure and edge systems.
Live speech translation supports more than 90 output languages and configurable translation timing, available in alpha for ElevenAPI enterprise customers.
Generally available multimodal analytics connects Cloud Storage files with Gemini through SQL, adding personal IAM authentication and array processing.
Generally available GPU instances protect model weights and inference data with hardware encryption and remote attestation across 36 zones.
Enterprise public beta adds continuous repository scanning, fix tracking, editable threat models, and API, webhook, and MCP integrations; administrator enablement is required.
Version 0.5.0 adds investigation time allowances, host-provided finding reviews, and GitHub Action support for additional inference providers.
Speech models and a voice agent SDK support multilingual conversations and tool calling through hosted APIs; private deployment requires contacting Mistral.
Personal Intelligence starts rolling out on Android Auto to eligible adult personal accounts in supported markets; Google built-in support follows in coming months.
Mini and Max speech-to-speech models support over 90 languages through ElevenAgents and a WebSocket API, with gradual workspace rollout and model-specific compliance and residency restrictions.
Locked chats gradually reach Free, Go, Plus, and Pro users across web, mobile, and desktop, hiding conversations behind PIN or supported device authentication.
Downloadable multilingual embedding model supports searching document text and page images containing charts, tables, and infographics.
Downloadable multilingual reranker scores text and document images for relevance, helping retrieval applications prioritize evidence containing charts, tables, and infographics.
Downloadable weights support self-hosted multimodal coding and agent workflows, with deployment guidance and up to one million tokens of context under the MiniMax Community license.
Diffusion language model with a million-token context window and hosted API access for conversational applications, tutoring, search, and agents.
Max subscribers gain Ultrafast options on web at higher credit usage; Mac app version 26.38 adds local folders to Projects.
Downloadable protein-design model generates sequences and structures using molecular constraints, with search and scoring to prioritize experimental candidates.
The 1017 update is available through API, web, and app, retaining the deepseek-flash API identifier, calling method, and pricing.
Authenticated APIs provide multilingual web search and page retrieval for agents and retrieval applications, with separate account limits of five requests per second.
Cloud agents create interactive canvases with restricted team sharing; environments support 300 repositories, and programmatic environment management is available in beta.
All API customers can create reusable voices from text descriptions for GPT-Live; voice creation is free, while generated speech incurs regular session charges.
The modernization agent converts supported Azure .NET service code to AWS equivalents in eight regions; Cosmos DB conversion remains in preview.
Early-beta terminal coding agent supports resumable conversations, parallel workspaces, tool approvals, MCP connections, and Python extensions.
Downloadable English reasoning model supports adjustable inference loops and up to 256K context; documented vLLM and llama.cpp integrations currently require Dory-specific builds.
Multimodal coding and agent model with downloadable weights, a million-token context window, and hosted inference options under a custom license.
Open-weight vision-language model for agents handling coding, tool use, webpages, charts, and documents, available through downloads and NVIDIA NIM.
Quantized agent model designed to fit in 24GB of memory for local coding, research, and tool-driven tasks on compatible NVIDIA hardware.
Four downloadable forecasting models support multivariate time series, probabilistic outputs, and CPU inference, with customization through Mistral Forge.
AI vulnerability scanning reaches general availability for public repositories and licensed private organization repositories; findings are advisory, consume AI Credits, and exclude fork and Dependabot pull requests.
Downloadable image generation and editing model supports ten reference images, bounding-box control, and API access, with commercial use governed by its license.
Instant Soundtracks generates music and synchronized sound effects for uploaded videos, with separate editable layers and existing-audio references through supported interfaces and API.
Versions 0.25.1–0.25.2 add connected project views, per-run MCP controls, and experimental collaboration with installed Claude Code or Codex; collaboration is off by default.
Collaborative specifications connect Markdown, designs, diagrams, and coding agents through MCP; open beta is rolling out to paid plans.
Speech-to-speech model is available through the Voice Agent API at unchanged pricing; applications using the latest-model alias upgrade automatically.
Document review agent is available worldwide with two free weekly reviews, PDF and plain-text inputs, and higher usage through a Gemini API key.
Generally available Gemini API reasoning model for complex coding, scientific work, and agents, accepting multimodal inputs and producing text.
Gemini 4 Argon rolls out globally to Google AI Pro and Ultra subscribers; macOS Antigravity coding is available in English for eligible adults in supported regions.
Swift library unifies local and hosted model interfaces; version 1.0 introduces breaking changes to JSON values and generated enums.