Google Gemma 4 Fires Open Shot.

TL;DR

Gemma 4 Launch: Google’s most advanced open model released, featuring multimodal capabilities and 140+ languages, explicitly challenging China’s open-source AI dominance.
Microsoft’s MAI Models: Microsoft Foundry introduces three in-house AI models, as Mustafa Suleyman projects full AI independence from OpenAI within two to three years.
Claude Code Vulnerability: A critical security flaw in Anthropic's leaked codebase enables credential exfiltration via prompt injection, triggering Congressional concern.
Midterm AI Funding: AI-linked campaign contributions exceed $185 million for the 2026 midterms, with campaign ads strategically avoiding direct mention of AI.
Anthropic Appeal: The DOJ formally appeals Judge Lin's Anthropic injunction to the Ninth Circuit, with a briefing deadline set for April 30.

Lead Story: Google Ships Gemma 4 — Reshaping the Open-Source AI Landscape

Google released Gemma 4 on Wednesday, its most capable open model to date, built on the Gemini 3 architecture. This launch is a direct strategic response to China's formidable presence in open-weight AI via DeepSeek, Qwen, and GLM-5.

The release comprises four distinct models: a 31B Dense model, currently ranked #3 on Arena AI's text leaderboard; a 26B Mixture of Experts variant; and two edge-optimized models (E4B and E2B) designed for near-zero latency operation on devices such as phones, Raspberry Pi, and Jetson Nano.

All models are distributed under an Apache 2.0 license, devoid of commercial restrictions or usage caps. They natively process images and video; the edge variants further extend capabilities with integrated audio input for speech recognition.

The multilingual scope is substantial, supporting over 140 languages out-of-the-box, with the larger models offering context windows up to 256K. Google asserts that its edge models outperform competitors 20 times their size on conventional benchmarks.

Gemma 4 is accessible across Hugging Face, Kaggle, Ollama, Google AI Studio, and as an Android AICore Developer Preview. NVIDIA has already optimized these models for local RTX inference.

The competitive landscape is rapidly intensifying. Meta's Avocado model faces delays until May or June following internal underperformance against Gemini. Concurrently, Alibaba launched Qwen 3.6-Plus this week, featuring a 1M-token context window and agentic coding capabilities that benchmark against Claude Opus 4.5. While Trending Topics noted Gemma 4's continued lag against certain Chinese competitors in raw benchmarks, its open licensing and on-device utility may represent a more significant market differentiator than leaderboard position alone.

In Other News

Microsoft launches three in-house AI models, accelerating its split from OpenAI. MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2 are now available on Microsoft Foundry. These offer transcription across 25 languages at 2.5x real-time speed, 60-second audio generation in under a second, and an image model debuting at #3 on Arena AI. VentureBeat characterized this as a "direct shot at OpenAI and Google." CEO Mustafa Suleyman has outlined a two-to-three-year roadmap for achieving complete AI independence from OpenAI, facilitated by an October 2025 agreement that permits Microsoft to develop its own frontier models while retaining OpenAI license rights through 2032.

Critical vulnerability found in Claude Code's leaked source. SecurityWeek reports a flaw that allows attackers to circumvent all deny rules via prompt injection. A maliciously crafted `CLAUDE.md` file could construct a 50+ subcommand pipeline, exfiltrating SSH keys, AWS credentials, and tokens without triggering security alerts. Users who installed Claude Code via npm between March 31, 00:21 and 03:29 UTC, may have inadvertently pulled a trojanized HTTP client. Rep. Josh Gottheimer has pressed Anthropic in Congress, warning that the leak exacerbates risks from CCP-backed exploitation attempts.

AI industry goes all-in on the midterms. ABC News reports that over $185 million in AI-linked contributions have already been deployed for 2026 races. Innovation Council Action (Sacks/Budowich) has pledged $100 million, while Brockman's Leading the Future PAC has raised $50 million. NBC News found that these AI-funded political advertisements saturate airwaves, yet conspicuously avoid any mention of AI itself, instead focusing on partisan flashpoints.

EU Parliament delays AI Act's high-risk rules to 2027. The Digital Omnibus proposal pushes compliance deadlines from August 2026 to December 2027, with product-embedded systems deferred further to August 2028. The IAPP reports that the Commission also failed to meet its own deadline for publishing technical standards, indicating continued regulatory inertia.

Gemma 4's Apache 2.0 licensing dominated developer discourse, with the prevailing sentiment that "Google just made the open-source argument for the West." Skeptics, however, pointed to benchmark gaps relative to Qwen 3.6-Plus. Microsoft's MAI launch drew sharper commentary: developers questioned whether Suleyman's independence timeline was "credible or corporate posturing," while enterprise buyers viewed it as leverage for OpenAI pricing negotiations. The Claude Code vulnerability report sparked debate on whether the prompt injection bypass was inherent to agentic coding tools or specific to Anthropic's architectural implementation. The midterms spending story elicited cynical reactions across the political spectrum, frequently summarized as, "they'll spend $185M to make sure nobody regulates the thing that's cutting 15,000 jobs a month."

One to Watch

States are racing ahead of Congress and the EU on AI regulation. Tennessee Governor Bill Lee signed SB 1580, prohibiting AI systems from impersonating qualified mental health professionals. Nebraska's chatbot safety bill (LB 1185) appears poised for passage. Georgia has three AI bills advancing to the governor's desk before legislative adjournment on April 6. In a more significant move, Governor Newsom signed an executive order allowing California to decouple its AI procurement decisions from federal mandates — a direct response to the Pentagon's Anthropic blacklisting. Axios positions California as "the national testing ground for AI rules." With the federal framework stalled and the EU delaying its own, states are effectively becoming the primary regulators, establishing precisely the fragmented regulatory landscape the White House aimed to preempt.

Quick Hits

Ninth Circuit docketed: The DOJ's appeal of Judge Lin's Anthropic injunction now holds an April 30 briefing deadline; an emergency stay motion is anticipated within days.
OpenAI acquires TBPN: OpenAI's first media acquisition is the successful tech talk show hosted by John Coogan and Jordi Hays, projected to exceed $30M in revenue this year, now housed under chief political operative Chris Lehane.
Q1 venture funding hit $300B globally: 80% of this capital went to AI, with OpenAI ($122B), Anthropic ($30B), xAI ($20B), and Waymo ($16B) collectively representing 65% of all global VC.
Alibaba shipped Qwen 3.6-Plus: This release, featuring a 1M-token context and agentic coding, marks Alibaba's third proprietary model drop in days, sustaining competitive pressure on Western open-source initiatives.
72% of Global 2000 companies now run AI agents beyond pilot phase, with Gartner projecting 40% of enterprise applications will embed task-specific agents by year-end.

Google's Gemma 4 and Microsoft's MAI models both launched this week — one reshaping open-source licensing, the other redrawing strategic partnerships — while states from Tennessee to California aggressively fill the regulatory void that Congress and the EU continue to widen. The interplay of these developments indicates a market in constant re-architecting, driven by intense competition and disparate governance approaches.