Sabi introduces RLNF to post-train generative AI on EEG brain signals

Released on 29 October 2026 under the Mistral Ecosystem Open-Weight licence, Mistral Large 4 permits unrestricted self-hosting, fine-tuning, and internal enterprise deployment, while commercial Model-as-a-Service redistribution requires a separate licence from Mistral AI. Trained in Europe, the updated weights incorporate additional reinforcement learning post-training for multi-repository software engineering and defensive cybersecurity workflows, and expose three selectable reasoning effort modes: Low, Medium, and High (the default).

As illustrated in the Hugging Face release architecture diagram, the model pairs a native Pixtral vision encoder with 72 sparse Mixture-of-Experts decoder layers supporting roughly a 1 million token context window. Mistral AI renders its benchmark comparisons interactively in HTML on the announcement page rather than as static chart images, reporting first place among open models on Cognition’s FrontierCode 2.0, a tie with GLM-5.3 on the Artificial Analysis Coding Agent Index using nearly half as many tokens, and 30% lower output token usage on Terminal-Bench 4.0 than its preview checkpoint.

Key numbers

• 19.7% on Cognition FrontierCode 2.0, ranking first among open-weight models.
• 48% on the Artificial Analysis Coding Agent Index while consuming 47% fewer tokens than GLM-5.3.
• 41% on Terminal-Bench 4.0, a 13 percentage point gain over the Mistral Large 4 preview checkpoint.

Jules operated as a background cloud agent that cloned GitHub repositories into isolated virtual machines to resolve issues, write tests, and open pull requests asynchronously. In its turndown notice, Google stated that architectural lessons and background execution patterns from Jules are being unified directly into Google Antigravity rather than maintained as a standalone web product.

Active Jules users can finish running tasks until 11 November 2026, when new task creation stops and workspaces enter read-only mode ahead of full deletion after 31 December 2026. To preserve project state, Google shipped an export command in the Jules Tools CLI that downloads account settings, environment variables, scheduled tasks, connected repository metadata, repository memories, environment setup scripts, and complete session logs and diff patches to local folders. Google has not announced automated cloud migration of saved Jules memories into Antigravity.

Previously, revising structured files produced by Gemini required regenerating whole drafts in chat or exporting files into Google Workspace editors first. With the new inline editor, users can prompt Gemini to turn raw inputs, such as photographed handwritten notes, directly into formatted PDF study guides, slide decks, spreadsheets, or text documents and preview the rendered pages side by side with the conversation.

Inside the preview pane, users can click directly into cells, slides, or paragraphs to make manual text edits, or highlight a specific text passage or visual element and ask Gemini to revise only that selection while leaving surrounding pages and formatting untouched. Multi-turn follow-up prompts update the live file in place. Google announced the capability on the official Gemini App X account and has not published per-plan file size caps or rollout restrictions by region.

The quota expansion doubles the volume of model turns, deep reasoning passes, and agentic tool executions available across OpenAI’s paid professional tiers: ChatGPT Pro 100, Pro 200, Pro 500, and Business Premium. Subscribers receive the expanded allowance automatically without plan migration or billing changes.

Addressing developer concerns around shared quota pools, Sottiaux confirmed that Pro 200 baseline limits were not trimmed beforehand, preserving existing capacity for ChatGPT Work and Codex software engineering sessions before the doubling took effect. Extended usage pools for OpenAI’s dot-series models also remain metered independently from default plan allowances rather than drawing down standard conversation quotas. OpenAI did not publish exact numeric message or compute-unit ceilings per billing window.

Instead of configuring Model Context Protocol (MCP) endpoints, system prompts, and helper scripts piecemeal per repository, developers can browse and enable curated Google and partner bundles from the Customizations panel in the Antigravity app or from the terminal via agy plugin install and agy plugin list. Every package is defined by a plugin.json manifest that versions and distributes up to five primitives together, alongside local schema checks through agy plugin validate.

Five bundled primitives

Each plugin combines progressive-disclosure Skills (Markdown instructions loaded on demand so idle tools do not consume context tokens), pre-configured MCP server connections, persistent workspace Rules, specialist Custom Agents, and deterministic lifecycle Hooks triggered at SessionStart, PreToolUse, PostToolUse, PreInvocation, PostInvocation, and Stop. Twelve third-party launch partners shipped verified plugins on 29 October 2026: Atlassian, Canva, CrowdStrike, Databricks, Datadog, Dropbox, Figma, monday.com, Ping Identity, SentinelOne, Stripe, and Unity, alongside Google bundles for Gemini API, Android CLI, Firebase, Chrome DevTools, and Google Workspace.

Cascaded voice stacks transcribe caller audio to text before an LLM can reason or invoke APIs, and existing speech-to-speech models typically wait until a speaker pauses before running external tools. Phonic-1 processes raw audio natively in full duplex, firing backend tool calls asynchronously while the caller is still talking so lookup results are already in context when the caller finishes speaking.

Across independent and enterprise voice agent evaluations shown below, Phonic-1 takes the top overall score on the Artificial Analysis Speech to Speech Index ahead of Grok Voice 3, GPT-Live-1, and Gemini 3.8 Live, while leading Sierra’s tau3-Voice retail, airline, and telecom suite and Cekura S2S at sub-second audio response latency.

Phonic-1 is available through Phonic’s developer API across more than 50 languages, accompanied by $27.5M in Series A financing led by First Round Capital with participation from Lux Capital, Felicis, and Definition, bringing total capital raised to $31.5M. Phonic has not released open model weights.

Earlier text-to-audio generators struggled when a single prompt described several consecutive events, often blending sounds simultaneously or dropping late cues. SFX v3 tracks ordered multi-part acoustic descriptions within one generation, such as rendering a crowd cheering, a ball dropping, and the crowd shifting to boos in chronological sequence without manual timeline splicing.

Acoustic fidelity also improves on sustained metallic and harmonic transients, cutting background hiss and ringing artefacts on gongs, chimes, and whistles compared with SFX v2 while strengthening short musical stingers and crowd or character vocal textures embedded inside effects. SFX v3 is live immediately in ElevenCreative, Studio, Flows, the ElevenLabs mobile app, and the sound-generation endpoint in ElevenAPI. ElevenLabs did not announce any change to per-generation credit pricing.

Developed by Canva’s 160-person research organisation from an open-weight foundation model post-trained with reinforcement learning on designer-curated layout and typography rewards, Brush focuses on modifying finished or in-progress designs rather than generating flat images from scratch. When selected inside the Canva AI model menu, Brush inspects the active document, writes and executes code to adjust targeted text, colours, spacing, or visual effects, and visually verifies the updated canvas while keeping untouched layers intact.

Across internal multi-page presentation, poster, and social-asset editing benchmarks, Canva reports sharp drops in both turnaround latency and inference cost compared with its previous general-purpose design agent architecture. Brush is rolling out in beta inside Canva AI for Canva Pro subscribers; Canva has not announced enterprise API access or open-weight plans.

Selecting the XR Blocks starter inside Google AI Studio Build prompts Gemini to generate browser-native spatial applications atop Google’s open-source three.js WebXR framework, bypassing native Unity or Unreal SDK setup. Generated projects automatically wire up bare-hand tracking, pinch gestures, depth-mesh physics collisions so virtual objects rest on real-world tables and walls, and Gemini-generated 360-degree 3D holograms built from text descriptions or uploaded photos.

Applications run with zero installation in Chrome version 136 or newer across Android XR headsets and glasses including Samsung Galaxy XR and XREAL AURA. Developers without spatial hardware can test and iterate on camera movement, hand pointers, and scene physics directly in Google AI Studio using XR Blocks’ built-in keyboard-and-mouse desktop simulator before sharing live links.