Gemma 4 12B open-source, Ideogram 4.0 open-weights, Perplexity Computer on Windows, Stargate Michigan

June 3, 2026 brings a wave of major open-source announcements: Google releases Gemma 4 12B (multimodal, edge, Apache 2.0) and launches Co-Scientist, its multi-agent system for scientific research. Ideogram opens up the weights for version 4.0. Perplexity brings its Personal Computer agent to Windows. OpenAI breaks ground on The Barn, a 1 GW datacenter in Michigan. GitHub Copilot rolls out a massive update across VS Code, JetBrains, Eclipse, and billing. On the xAI side, Grok Imagine 1.5 Preview is available via API and Grok voices arrive on Vapi.

Gemma 4 12B — open-source multimodal model for laptops

June 3 — Google launches Gemma 4 12B, a unified multimodal model designed to run directly on laptops and edge devices (edge computing). Its main technical distinction is its encoder-free architecture, which simplifies deployment and reduces software dependencies. The model is released under the Apache 2.0 license, making it freely usable for commercial and research purposes.

Feature	Value
Parameters	12 billion
Modalities	Text + Image (unified)
Architecture	Encoder-free (encoder-free)
License	Apache 2.0
Hardware target	Laptops, edge computing
Availability	Downloadable now

The announcement generated 620,000 views and 5,000 likes on X, making it one of the week’s most engaging open-source announcements. Reposted by Google DeepMind, Gemma 4 12B continues the Gemma line of models, but takes a step forward by combining vision and text in a single compact model.

For developers, this model opens a concrete path to deployable multimodal AI applications outside the cloud — on machines with consumer GPUs, without an API subscription.

🔗 Launch tweet @googlegemma

Google DeepMind Co-Scientist — multi-agent partner for scientific research

June 2 — Google DeepMind introduces Co-Scientist, a Gemini-based multi-agent system designed to act as a true partner in scientific research. Its core capability: generate, debate, and evolve hypotheses on complex scientific problems autonomously, with minimal human intervention.

Co-Scientist is not just a summary or document-search tool. It can propose original hypotheses, internally challenge them between agents, then refine them — an ideation-debate-evolution cycle previously reserved for human research teams. The system has been evaluated with global scientific experts on specialized benchmarks.

Availability: early access for individual researchers via Hypothesis Generation (waitlist). Collaborations with global scientific experts are underway to evaluate performance in specific domains.

“We believe AI can be a dedicated research partner to help discover the next breakthrough.” — @GoogleDeepMind

The announcement generated 144,000 views on X. In the context of a scientific sector under pressure to accelerate discovery cycles, Co-Scientist targets individual researchers directly who do not have the resources of a large team.

🔗 DeepMind article

Ideogram 4.0 — downloadable open-source weights, fine-tuning possible

June 3 — Ideogram announces Ideogram 4.0, presented as “the best open-source image model in the world.” The announcement stands out for its radically open philosophy: the model weights are downloadable, developers can fine-tune it on their own data, and they can deploy it on their own infrastructure.

Aspect	Detail
Availability	All Ideogram plans + API (now)
Hugging Face weights	`ideogram-ai/ideogram-4-nf4`
Demo	huggingface.co/spaces/multimodalart/ideogram4
Fine-tuning	Allowed on proprietary data
Deployment	Personal infrastructure possible
X engagement	396,800 views — strong virality

The announcement was echoed and endorsed by Hugging Face as an example of the progress in open model quality. The slogan “Think it. Make it. Own it.” sums up the philosophy: unlike typically closed image generation models (Midjourney, DALL-E, recent Stable Diffusion), Ideogram 4.0 offers a full-ownership model for studios and developers.

For creative teams, the appeal is immediate: customize style on their data without dependence on an external service, and deploy internally.

🔗 @ideogram_ai announcement

Perplexity Personal Computer arrives on Windows

June 3 — Perplexity announces the arrival of its Personal Computer on Windows agent. Previously available only on macOS, the agent is about to conquer Microsoft’s operating system.

“Personal Computer is coming to Windows. Personal Computer for Windows runs on your machine and orchestrates across the apps and files you use every day. We’ll roll out first to paying Max and Enterprise Max subscribers on the waitlist.” — @perplexity_ai

Personal Computer runs locally on the user’s machine — it orchestrates everyday applications and files without going fully through the cloud. This hybrid local/cloud approach places Perplexity in a distinct category from purely cloud-based agents: the user retains control over sensitive data.

Deployment will begin with Max and Enterprise Max subscribers on the waitlist via perplexity.ai/personal-computer.

Perplexity Computer — 400+ integrations for businesses

On the same day, Perplexity clarifies the scope of the agent’s enterprise capabilities: more than 400 tools connected across all types of businesses. The cited integrations include Intuit QuickBooks (accounting), Vercel (web deployment), Shopify (e-commerce), and Canva (design).

This complementary announcement positions Perplexity Computer as a universal enterprise agent capable of integrating into existing workflows without replacing current tools.

🔗 Enterprise integrations

Stargate Michigan — The Barn, 1 GW datacenter in Saline

June 1 — OpenAI and Governor Gretchen Whitmer break ground on The Barn, a 1 gigawatt datacenter campus in Saline, Michigan. The project is part of OpenAI’s Stargate program — a long-term infrastructure-building effort for advanced AI, in partnership with Oracle, Related Digital, Walbridge, and Blackstone.

Metric	Value
Capacity	1 GW
Location	Saline, Michigan
Construction jobs	2,500+ (union)
Permanent jobs	450
County + indirect jobs	2,500
Rec center investment	$10M (Oracle + Related Digital)
Projected tax revenue	$1 billion (lease term)
Student Codex credits	up to $45M
Eligible students	400,000+

Community commitments are explicit: local residents will not bear infrastructure costs on their electricity bills. Cooling operates in a closed loop, with water usage equivalent to an ordinary office building.

The Michigan student program offers up to $45 million in Codex credits to the 400,000 students at the state’s universities, community colleges, and trade schools (18 and over, 2026-2027 academic year).

🔗 openai.com announcement

GitHub Copilot in VS Code — May releases (v1.120–v1.123)

June 3 — GitHub publishes the monthly Copilot recap for VS Code, covering versions v1.120 through v1.123 delivered throughout May 2026. This broad update consolidates several months of work on agents, synchronization, and security.

The structural updates:

Agents window (stable preview): agent-first interface for working on multiple projects simultaneously, with quick navigation and change review
Remote agents (preview): run on remote machines via SSH or Dev Tunnels — sessions persist even if the client disconnects
Agent Host Protocol (AHP): open protocol for synchronizing session state across multiple clients
Session sync: chat sessions automatically sync to the GitHub account, accessible across machines and workspaces
/chronicle: query past sessions, generate daily stand-up reports, get personalized productivity tips
Terminal security: passwords, PINs, and verification codes entered in the terminal are never sent to the language model
Command risk assessment (experimental): terminal confirmations include AI-generated risk levels
BYOK air-gapped: BYOK models work in isolated environments without GitHub access
Configurable utility models: model choices for titles, summaries, rename suggestions, commit messages, and intent detection
Integrated browser: mobile device emulation, advanced screenshots (viewport, selected area, full page), local HTML preview
Mermaid rendering and YAML front matter: natively integrated into Markdown preview
Terminal output compression: verbose patterns (tests, builds, linters, Docker, package managers) automatically compressed

The multi-agent work model is taking shape: multiple agent sessions can run in parallel in the Agents window, each with its own context and its history synchronized to the GitHub account.

🔗 VS Code May changelog

Copilot Code Review — MCP skills and Medium analysis level (public preview)

June 2 — GitHub launches two major features for Copilot code review in public preview, making it possible to adapt automated reviews to each team’s context and complexity.

Skills and MCP support: teams can now configure custom agent skills (.github/skills/code-review/SKILL.md files) and connections to MCP servers to enrich each review with internal tool context — issue tracking, documentation, service catalogs, incident tools. MCP configuration is done from repository settings → Copilot → MCP servers. Any existing MCP configuration for the cloud agent automatically applies to reviews — one-time investment, consistent behavior across both agents.

“Medium” level: a new intermediate tier routes complex pull requests to a higher-reasoning model, designed to detect subtle bugs, analyze security logic, and handle cross-service changes. The “Low” level remains the fast, economical default for simple changes. Medium consumes more AI Credits than Low, with clear cost signals in the interface. Admins define the level per repository.

Actions workflows are also configurable, making it possible to control the compute environment used by Copilot for reviews. Available for Copilot Pro, Pro+, Business, and Enterprise.

🔗 Shape Copilot Code Review

Copilot CLI and stronger agentic capabilities in JetBrains IDEs

June 2 — GitHub announces the arrival of Copilot CLI in JetBrains with an agent selector that lets users choose between Agent mode (autonomous), Ask (quick help), Custom agents, or Plan (collaborative planning). Three new commands enrich CLI sessions:

/remote : control a CLI session from github.com or the mobile app
/compact : manually compress the context of long sessions
/chronicle : analyze history, generate stand-up reports, and get personalized advice

The agent debugging panel (public preview) displays a chronological log of interactions — particularly useful for debugging custom agents and complex multi-agent workflows. The reasoning effort of reasoning models is configurable directly from the model selector, allowing the depth of analysis to be adjusted according to the task.

Cloud agent sessions are now integrated into the unified sessions view, making it possible to manage all sessions (local, CLI, cloud) from one place. A custom agent editor centralizes the creation and management of skills, instructions, and prompts.

New sign-in options: Google and Apple are added to the available sign-in methods. Among the features that have reached GA: agent skills, agent hooks, prompt files, and Anthropic Thinking. A gradual rollout is underway to make Copilot CLI agent the default experience in JetBrains.

🔗 Copilot JetBrains changelog

GitHub Copilot billing and plans — usage-based billing GA, Copilot Max

June 1 — GitHub activates usage-based billing (AI Credits) for all Copilot plans as of June 1, 2026. Each plan includes a monthly AI Credits quota; beyond that, users can continue by configuring an additional spending budget with limits defined by admins.

Structural changes:

Copilot code review now consumes GitHub Actions minutes in addition to AI Credits — admins can configure a default runner for all reviews at the organization level, without per-repository configuration
Per-user budgets in GA: organization and enterprise admins define universal or per-user-specific budgets, with email notifications as limits approach
Copilot Max: upgrade available today for existing Student, Pro, and Pro+ subscribers — higher quota and higher spending limits for intensive workflows
New sign-ups suspended for Student, Pro, Pro+, and Max — reopening expected in the coming weeks

A note for Enterprise teams: admins will need to enable the alternative GPT-5.5 model via model policies to replace the deprecated GPT-4.1. Once enabled, GPT-5.5 appears in the Copilot Chat model selector in VS Code and on github.com.

🔗 Billing and plans update

Grok Imagine 1.5 Preview available in the xAI API

June 3 — xAI announces the availability of Grok Imagine 1.5 Preview in the xAI API, accessible at x.ai/api/imagine. This is a major update to Grok’s image generation capability — version 1.5 Preview succeeds Grok Imagine and represents a new generation of xAI’s image model, available directly via API for developers.

Aspect	Detail
Model	Grok Imagine 1.5 Preview
Access	xAI API — x.ai/api/imagine
Status	Preview (no official product page at the time of the scan)
Availability	Developers via API now
X engagement	164,800 views, ~1,000 likes

The announcement tweet came from the @grok account and was immediately reposted by @xai — a signal of priority for the team. Grok Imagine was already accessible in the Grok interface, but API availability for version 1.5 Preview opens xAI image generation to developers looking to integrate it into their applications.

Grok Imagine 1.5 Preview is part of xAI’s growing momentum in the multimodal segment: the same day, the Grok STT and Grok TTS voice APIs expand their reach via Vapi integration. The xAI ecosystem now covers text, image, and voice via API, offering a full alternative to OpenAI and Google offerings in this segment.

🔗 Announcement @grok

Grok STT and TTS available on Vapi

June 3 — xAI voice APIs — Grok STT (speech recognition) and Grok TTS (speech synthesis) — are now integrated into Vapi, an enterprise voice AI platform. Launched in April 2026, these APIs now extend their reach to enterprise voice agent developers via Vapi.

The announcement generated 433,300 views, 1,100 likes, and 272 reposts — strong engagement for a technical integration announcement. Vapi makes it possible to create custom voice agents; the Grok STT/TTS integration adds xAI voice models there as a competing option to those from ElevenLabs, Deepgram, or OpenAI.

🔗 Grok on Vapi

NVIDIA CVPR 2026 — Physical AI agent skills for AV, robotics, AI vision, and Alpamayo 2 Super 32B

June 3 — On the occasion of CVPR 2026 (Denver, June 3–7), NVIDIA unveils a new generation of Physical AI agent skills covering the full research cycle for autonomous vehicles, robotics, and AI vision.

Key announcements for autonomous vehicles:

AlpaGym: closed-loop reinforcement learning framework, scalable to thousands of GPUs
OmniDreams: action-conditioned generative world model
Alpamayo 2 Super: open-source 32-billion-parameter VLA (Vision-Language-Action) model for level 4 autonomous driving

For robotics: Isaac Sim 6.0 with native agent connectors, mobility skills, sim-to-real reinforcement learning, and Cosmos-H-Surgical-Simulator for surgical robotics.

For AI vision: Metropolis skills to generate synthetic visual anomalies and the Defect Image Generation skill for industrial inspection.

The NVIDIA Physical AI Dataset exceeds 15 million downloads on Hugging Face. The skills are available on GitHub and NVIDIA Brev with free trial credits on H100.

🔗 NVIDIA CVPR 2026 Physical AI

NVIDIA Transaction Foundation Models — Revolut, Mastercard, Adyen, Stripe

June 1 — NVIDIA explains how major financial institutions are converging on transaction foundation models — foundation models trained on billions of proprietary financial events.

Institution	Approach	Key result
Revolut	PRAGMA — 24 billion events, 26M users, 100+ countries	A single model outperforms specialized models (fraud, credit, recommendations)
Mastercard	Large tabular model on billions of transactions (NVIDIA NeMo AutoModel)	Applications: fraud, credit, loyalty
Adyen	Models deployed at the scale of $1 trillion in payments, RL for conversion	”0.1% authorization = massive impact”
Stripe	Transactional behavioral context	$112 billion in fraud blocked/year, -38% fraud rate

NVIDIA provides a development example, “Build Your Own Transaction Foundation Model,” available on AWS SageMaker HyperPod and Nebius AI Cloud.

🔗 NVIDIA Transaction Foundation Models

Anthropic — Partner Network, cyber threats, and Claude Code /fork

Claude Partner Network — Services Track and Partner Hub

June 3 — Anthropic expands its partner network with two new structures: the Services Track and the Claude Partner Hub.

The Services Track is a three-tier system evaluating firms on their real-world practice with Claude:

Tier	Active certified	Deployed clients	Public testimonials
Select	10+	2+ (12 months)	1+
Preferred	100+	15+	3+
Global Premier	1,000+	100+ (3 regions)	15+

The Claude Partner Hub is a portal with daily updates on each partner’s progress, and an MCP connector to query Hub information directly from Claude.

Key figures: $100 million invested in partner training, 40,000+ applications since launch in March, 10,000+ certified consultants to date. Among the participating firms: Accenture (30,000 professionals trained), Deloitte (470,000 people), KPMG (276,000+), Cognizant, PwC.

🔗 Services Track and Partner Hub

MITRE ATT&CK report — one year of cyber threats activated by AI

June 3 — Anthropic’s Frontier Red Team publishes an analysis of 832 banned accounts between March 2025 and March 2026, mapped to the MITRE ATT&CK framework.

Three main conclusions:

Indicator	Figure
Accounts analyzed	832 (March 2025 – March 2026)
AI used for malware writing	560/832 (67.3%)
Medium+ risk actors at beginning of period	33%
Medium+ risk actors at end of period	56% (+1.7×)
Espionage operation dismantled	November 2025 — max score 100/100

Agentic orchestration — a model that autonomously chains together the steps of an attack — is not yet covered by the MITRE ATT&CK framework. Anthropic has begun discussions with MITRE to evolve the framework, and is publishing these data in the wake of Verizon’s DBIR 2026 report.

🔗 Anthropic report — AI cyber threats

Claude Code /fork revamped — background agent with exact context

June 3 — Anthropic changes the behavior of /fork in Claude Code. The new /fork launches a background agent that inherits the exact context of the current session (system prompt, tools, history, model, prompt cache) and returns its result to the original session.

Command	Behavior
`/fork` (new)	Launches a background agent with exact context + cached prompt → result returned in the session
`/branch` (ex-/fork)	Copies the transcript to a new manually driven session

This change places /fork in Claude Code’s evolution toward a multi-agent work model.

🔗 Tweet @ClaudeDevs

OpenAI Blueprint — democratic governance of frontier AI

June 3 — OpenAI publishes a public policy document (“blueprint”) describing how the United States can build a durable federal framework to govern frontier AI systems.

3-part strategy:

National framework — build on the emerging consensus from state laws (California SB 53, New York RAISE Act, Illinois SB 315)
Strengthen CAISI — make the Center for AI Safety and Innovation the leading federal institution for frontier AI safety
National resilience plan — mobilize the government around the national security and public safety challenges posed by frontier AI

The blueprint follows the “Promoting Advanced Artificial Intelligence Innovation and Security” Executive Order published the day before by the White House.

🔗 OpenAI Blueprint

Gemini CLI v0.45.0 and Wan creative tools

Gemini CLI v0.45.0

June 3 — Google releases the stable v0.45.0 version of Gemini CLI with an architectural overhaul of ContextManager (the core component for managing conversational context) to improve robustness and performance.

Feature	Detail
Context Simplification	Major redesign of ContextManager
A2A Usage Metadata	Exposure of metadata in the Agent-to-Agent protocol
Termux fixes	Resolution of infinite retry loops (Android environment)
PTY resize	Fix for resize errors
Topic updates	Forced sequential execution to avoid conflicts

🔗 Gemini CLI changelog

Wan — new Extract Line Art and Render skills

June 3 — Wan (Alibaba Tongyi Lab’s video and image model) launches two new creative tools: Extract Line Art (conversion of complex visuals into clean, precise line drawings) and Render (instant transformation of line drawings into finished visuals). Both skills are available on create.wan.video/lab/skill. The announcement generated nearly 600,000 views in a few hours.

🔗 Wan announcement

GPT-4.1 removed from Copilot — replaced by GPT-5.5

June 2 — GitHub officially deprecates GPT-4.1 across all GitHub Copilot experiences (chat, inline edits, ask and agent modes, code completions), effective June 1, 2026. The recommended alternative is GPT-5.5. Enterprise admins must enable access to GPT-5.5 through model policies in their Copilot settings. No manual action is required to remove GPT-4.1.

🔗 GPT-4.1 deprecated

Copilot Code Review for Azure Repos (technical preview)

June 2 — GitHub launches Copilot code review for Azure Repos in technical preview. Users can request a Copilot review directly from an Azure DevOps pull request — inline comments, improvement suggestions, issue detection. No GitHub Copilot license is required for Azure DevOps customers. Usage is billed in GitHub AI credits, independently of existing Copilot plans.

🔗 Copilot Code Review — Azure Repos

Cohere Co/plot — open-source visualization tool for AI research

June 3 — Cohere Labs releases co/plot, an open-source data visualization tool for AI researchers. Born from a concrete need during the development of Tiny Aya (a multilingual model for 70+ languages), co/plot addresses the limitations of Matplotlib (full reruns on every change) and Figma (disconnected from real data).

The tool offers rapid prototyping with a predefined and customizable style, grounded in real data. Published in the spirit of open science, it is available at coplot.vercel.app.

🔗 Cohere co/plot

Briefs

Claude Code v2.1.161 — New version: OTEL_RESOURCE_ATTRIBUTES as metrics labels, independent parallel tool calls in the event of partial failure, improved Linux clipboard (wl-copy/xclip/xsel), /autofix-pr fixes in git worktrees. 🔗 Releases
Anthropic supports the AI Executive Order — Anthropic welcomes the EO “Promoting Advanced Artificial Intelligence Innovation and Security” and announces its collaboration with the White House on its implementation. 🔗 AnthropicAI tweet
GitHub Copilot in Eclipse — BYOK available for Business and Enterprise, refreshed chat interface, skills and prompt files, Thinking blocks, better ABAP support. 🔗 Eclipse changelog
NVIDIA Research CVPR — NVIDIA Research presents work on advanced grasping, intelligent autonomous driving, and large-scale agent training at CVPR 2026. 🔗 NVIDIA Blog
Fun-Realtime-TTS #1 Speech Arena — Alibaba’s Fun-Realtime-TTS (TONGYI_SpeechAI team) reaches first place on Artificial Analysis’s Speech Arena Leaderboard (Elo 1,219), ahead of Google Gemini 3.1 Flash TTS and Inworld Realtime TTS-2. 🔗 Tweet @Ali_TongyiLab
Go by Gopuff — AI shopping assistant — Launch of “Go by Gopuff,” a personal shopping assistant co-developed with xAI, powered by Grok models (text, audio, image) for express delivery in just a few minutes. 🔗 xAI announcement
OpenAI teaser “It’s time to fly.” — OpenAI posted a pinned tweet “It’s time to fly.” with an embedded video (68k+ views, reposted by @OpenAIDevs). No associated product page was available at the time of the scan — an imminent announcement not yet published. 🔗 OpenAI tweet

What this means

The open-source/open-weights wave is gaining momentum. Gemma 4 12B (Apache 2.0, edge, multimodal) and Ideogram 4.0 (downloadable weights, unrestricted fine-tuning) embody a structural trend: major teams are releasing competitive models under permissive licenses, narrowing the gap between closed offerings and open-source alternatives. For developers, June 3 marks a rare moment when a leading vision model and a leading image model become simultaneously accessible without commercial constraints.

Local/desktop agents are becoming the new competitive front. Perplexity Personal Computer on Windows — which runs directly on the machine — responds to a clear demand: companies want the autonomy of agents without exposing sensitive data to the cloud. DeepMind’s Co-Scientist targets individual researchers with a scientific reasoning partner. Anthropic’s revamped Claude Code /fork as a multi-agent background agent completes the picture: three different ecosystems (Perplexity, Google, Anthropic) converge on the same paradigm — an agent that works in parallel, with rich context, without constant human intervention.

GitHub Copilot consolidates its lead in developer tooling. The June wave — usage-based billing GA, Copilot Max, Code Review Medium, JetBrains CLI, session sync, BYOK air-gapped — is not a set of isolated features but a coherent platform pushing toward a model: a single agent accessible from all IDEs, all machines, with a level of control (MCP, skills, reasoning effort) adapted to each team. The replacement of GPT-4.1 with GPT-5.5 fits into this logic of continuous upgrades.

Physical AI and AI infrastructure are taking root in the real economy. Stargate Michigan (1 GW, 2,500 union jobs, $1 billion in projected tax revenue) and NVIDIA CVPR 2026 (Alpamayo 2 Super 32B for level 4 autonomous driving, Physical AI datasets with 15 million downloads) signal that AI is moving beyond the purely software domain into high-physical-intensity areas — data centers, autonomous vehicles, robotics. Financial institutions (Revolut, Mastercard, Adyen, Stripe) follow the same logic: foundation models trained on their proprietary data, not generic APIs.