AI Product Updates Daily — June 4, 2026

What changed in AI products — June 3–4, 2026

Six things moved at once: Google shipped an open model that runs on a laptop without a separate vision encoder, Microsoft put its first in-house coding model inside GitHub Copilot, Meta flipped its customer-service AI agent to every business on the planet, the UK regulator handed publishers a legal escape hatch from AI Overviews, Trump signed a frontier model review order with no teeth, and the EU placed a €320B bet that it can build its own AI stack from scratch. Below is the full breakdown.

Google Gemma 4 12B: encoder-free, runs on 16GB

Google published Gemma 4 12B on June 3, a new 12B-parameter open-weight model that strips out the separate vision and audio encoders most multimodal models depend on.1 Instead, visual patches (48×48 pixels, projected by a 35M-parameter embedder) and raw audio frames (40ms slices at 16kHz, linearly projected) enter the same decoder weights as text — making Gemma 4 12B the first medium-sized Gemma model to handle audio natively.

The practical upshot: a single model fits in 16GB VRAM or unified memory, handles text, images, and audio in one forward pass, and can be fine-tuned with LoRA in a single training step without touching separate encoder parameters. Google is shipping a macOS desktop app (Google AI Edge Gallery) that runs it fully offline on Apple Silicon, and a CLI command (litert-lm serve) that exposes it as a drop-in OpenAI-compatible local server.2

On benchmarks, performance sits close to Gemma 4 26B while using roughly half the memory.3 Available now on Hugging Face and Kaggle under Apache 2.0.

Gemma 4 benchmark comparison across model sizes — Gemma 4 12B vs. Gemma 3 27B and Gemma 4 26B across eight benchmarks 3

Releasing entity: Google DeepMind. Date: June 3, 2026. Availability: GA, open weights.

Microsoft MAI-Code-1-Flash: in GitHub Copilot, beats Claude Haiku 4.5

Microsoft's Superintelligence team released MAI-Code-1-Flash on June 2 — a 137B MoE coding model built entirely on clean, licensed data with no OpenAI model distillation.4 It is now rolling out to Copilot Free, Pro, Pro+, and Max users in VS Code.5

Benchmark claims from Microsoft's own model card:

Benchmark	MAI-Code-1-Flash	Claude Haiku 4.5
SWE-Bench Pro	51.2%	35.2%
IF Bench (instruction following)	+28.9 pts lead	—
Advanced IF	+14.5 pts lead	—
Average token use	up to 60% fewer	baseline

Scatter plot comparing MAI-Code-1-Flash vs Claude Haiku 4.5 on pass rate and token usage — MAI-Code-1-Flash (green) vs. Claude Haiku 4.5 (orange) — higher pass rate, lower token use 4

The model adapts its reasoning budget — using more compute for complex tasks and staying lightweight for simple completions — which Microsoft says drives the token efficiency.4

This is the clearest signal yet that Microsoft intends to shift GitHub Copilot off third-party model dependence. MAI-Code-1-Flash is part of the broader seven-model MAI portfolio announced at Build 2026, which also includes MAI-Thinking-1 (the reasoning model). No additional setup for existing Copilot users; the model appears in the VS Code model selector.

Releasing entity: Microsoft Superintelligence. Date: June 2, 2026. Availability: Rolling out to all Copilot plans.

Meta Business Agent: globally available on WhatsApp, Instagram, Messenger

Meta opened Meta Business Agent to all businesses worldwide on June 3, following roughly two years of testing in India and Mexico.6 The agent handles customer questions, product recommendations, appointment booking, lead qualification, and human handoff across WhatsApp, Instagram DMs, and Messenger.7

More than 1 million businesses were already using the agent on WhatsApp and Messenger before today's global launch. Deployment takes "minutes" with no code required; a business enters its product catalog and tone, and the agent starts answering in the customer's local language. Meta is also testing a morning briefing feature that surfaces overnight conversation summaries for business owners.

Pricing: free to start, with paid tiers coming in "the coming months." Large enterprises will pay per token. Third-party integrations (Shopify, Zendesk, Shopee) are available through the Meta Business Agent Platform.

about.fb.comhttps://about.fb.com/news/2026/06/meta-business-agent/外部リンク

コンテンツカードを読み込んでいます…

Releasing entity: Meta. Date: June 3, 2026. Availability: GA globally.

Google Search Console: AI visibility reports now in limited beta

Google launched Search Generative AI performance reports in Search Console on June 3, giving site owners dedicated views of how their pages appear inside AI Overviews and AI Mode.8 The reports show impressions, top pages, countries, devices, and date-level granularity.

Google Search Console generative AI performance report interface — The new dedicated generative AI performance report in Search Console 8

The launch is limited to a subset of websites while Google gathers feedback. Data will also continue to appear in the existing overall performance report.

Releasing entity: Google Search. Date: June 3, 2026. Availability: Limited rollout.

Anthropic: Claude Partner Network adds service tiers ahead of IPO

Anthropic formalized its enterprise channel program on June 3, adding a tiered "Services Track" to the Claude Partner Network it launched in March.9 The program received more than 40,000 applications since March; the current active partner roster sits around 100 companies including Accenture and Cognizant.

New tier structure:

Select — at least 10 Anthropic-certified personnel
Global Premier — at least 1,000 Anthropic-certified personnel

A Partner Hub portal lets partners track certification status and connects clients with qualified firms. Anthropic committed $100M to partner training, technical support, and co-marketing when it launched the program.

The timing aligns with Anthropic's confidential S-1 filing (June 1, 2026) ahead of a planned IPO at a ~$965B valuation. The partner buildout is explicitly meant to demonstrate revenue durability to prospective public-market investors.10

Releasing entity: Anthropic. Date: June 3, 2026.

Regulation and policy

UK CMA orders Google to give publishers an AI opt-out — 9 months to comply

The UK Competition and Markets Authority issued conduct requirements on June 3 ordering Google to let UK publishers opt their content out of AI Overviews and out of being used to fine-tune Google's AI models, while remaining indexable in regular search.11 Google must also provide clearer data on how AI features affect site traffic and include attribution links in AI-generated summaries.

The CMA gave Google 9 months to implement all changes, though it expects publisher controls to arrive "well before" the deadline. Google must file compliance reports every 6 months. The ruling comes under the UK's Digital Markets, Competition and Consumers Act — the first binding regulation of AI search features globally.12

Trump signs AI executive order with 30-day voluntary review for frontier models

President Trump signed "Promoting Advanced Artificial Intelligence Innovation and Security" on June 2, creating a voluntary framework where AI developers can submit frontier models for up to 30 days of pre-release review by the NSA and CISA.13 The review focuses on national security and cybersecurity risks. Participation is entirely optional; there is no penalty for companies that decline to submit models.

The order also directs agencies to establish a 30-day AI cybersecurity information exchange with critical infrastructure operators, and requires CISA to push AI-driven defense tools to federal systems, rural hospitals, community banks, and utilities.13

A prior broader AI order was cancelled last month; this version deliberately avoids any mandatory pre-release approval requirement, leaving the definition of "covered frontier model" to be worked out over the next 60 days via classified agency process.

EU unveils €320B tech sovereignty plan

The European Commission announced a ten-year Tech Sovereignty Package on June 3 targeting cloud, AI, and semiconductors, with a €320B ambition funded 4:1 private-to-public.14 A €180M Sovereign Cloud Framework contract has already been awarded to four EU providers. The package includes open-source requirements for publicly funded AI projects and explicit intent to reduce EU government reliance on AWS, Azure, and GCP.

Also notable

Meta employee tracking partial rollback. Following Reuters reporting and staff pushback, Meta modified its Model Capability Initiative — a program recording keyboard and mouse activity for AI training — to allow employees 30-minute opt-out windows for personal tasks. There is no permanent opt-out option. The EU GDPR implications for non-US employees remain unresolved.14

Let's Encrypt moves to post-quantum certificates. The CA covering 300+ million websites began issuing hybrid post-quantum certificates combining ML-KEM (Kyber) with traditional ECDSA on June 3, with full PQC support targeted for 2028.14

Consumer DRAM prices surge. AI data center demand for high-bandwidth memory has pushed 32GB DDR5 to a minimum of $375 — up from $90 in early 2024 — as manufacturers redirect fab capacity to HBM production. DRAM contract prices rose 90–95% quarter-over-quarter in Q1 2026. Analysts don't expect relief before 2027.14

xAI pauses specialist Grok trainer hiring. Elon Musk's xAI temporarily halted hiring for specialized chatbot trainers while navigating staffing and organizational changes.15

Next scheduled issue: June 5, 2026