Skip to content
FAQ
AI & ML Analysis

AI Week, Mar 30–Apr 2, 2026: Gemma 4 Goes Apache 2.0 as OpenAI Raises $122B and Buys a Podcast

Google DeepMind released Gemma 4, 4 open models under Apache 2.0 that span phones to a single H100, while OpenAI disclosed a $122 billion round at an $852 billion valuation and acquired the tech podcast TBPN.

6 min read

Written on November 15, 2026, covering news from Mar 30–Apr 2, 2026.

Google DeepMind shipped Gemma 4, an open-model lineup released under Apache 2.0 terms and built to run on hardware ranging from phones to a single data-center GPU. OpenAI, for its part, disclosed a $122 billion funding round and bought the tech podcast TBPN.

Thursday, April 2

Google DeepMind released Gemma 4, a family of open models licensed under Apache 2.0. Google called it its most intelligent open model family so far and pitched it at advanced reasoning and agentic workflows. According to Google, the models rest on the same research and technology base as Gemini 3.

Four sizes, one family

Gemma 4 comes in 4 sizes: Effective 2B and Effective 4B for phones and edge hardware, a speed-oriented 26B Mixture of Experts model, and a 31B Dense model tuned for output quality. Google said the lineup is meant to cover Android devices and laptop GPUs at one end and developer workstations and accelerators at the other.

Per Decrypt, the E2B and E4B models can run fully offline, take audio input natively and offer a 128K context window. The 26B and 31B models extend context to 256K and add native function calling plus structured JSON output for agent building. All 4 models process images and video natively.

Decrypt reported that full-precision weights for the larger models fit on one 80GB NVIDIA H100 GPU, while quantized versions run on consumer hardware. Google said Gemma 4 was natively trained on more than 140 languages.

The license switch

Decrypt described the move to Apache 2.0 as a significant break from earlier Gemma releases, which used a custom license that it said created legal ambiguity for commercial products. Under the new terms, Decrypt noted, developers can modify, redistribute and commercialize the models without worrying that Google will change the rules later.

Hugging Face co-founder Clement Delangue welcomed the release, saying “Local AI is having its moment”. Google DeepMind CEO Demis Hassabis described Gemma 4 as “the best open models in the world for their respective sizes”.

Rankings and early testing

At launch, the 31B Dense model sat third among all open models on Arena AI’s text leaderboard, with the 26B MoE in sixth. Chinese models held the top 2 spots. Google claims both larger Gemma 4 models outperform models 20 times their size, and Decrypt said that claim holds up against the Arena AI figures. Decrypt also noted that closed systems from Anthropic, OpenAI and Google’s own Gemini remain ahead on the toughest benchmarks.

In Decrypt’s own hands-on test, the model applied reasoning even to simple prompts, which could make answers feel over-engineered, and its creative writing was rated serviceable rather than inspired. Code was the strongest showing: a game it generated ran without errors on the first attempt.

Decrypt placed the launch against the rise of Chinese open models, citing its own earlier reporting that their share of global open-model usage grew from barely 1.2% in late 2024 to roughly 30% by the end of 2025. Previous Gemma generations have been downloaded more than 400 million times, and the community has produced more than 100,000 variants.

On-device tooling

Alongside the models, Google opened access to Android’s built-in Gemma 4 through a new AICore Developer Preview. Its Google AI Edge Gallery app, available on iOS and Android, gained Agent Skills, which Google called one of the earliest apps able to carry out multi-step, autonomous agent workflows fully on the device.

For developers deploying beyond that app, Google pointed to LiteRT-LM, which it said can, thanks to GPU optimizations, handle 4,000 input tokens spread over 2 separate skills in less than 3 seconds. Google’s published edge figures: on a Raspberry Pi 5 running on CPU, 133 prefill and 7.6 decode tokens/s; with NPU acceleration on the Qualcomm Dragonwing IQ8, 3,700 prefill and 31 decode tokens/s.

Google also released a Python package and a litert-lm CLI for Linux, macOS and Raspberry Pi, which supports tool calling and lets developers try Gemma 4 without writing code. The 31B and 26B models can be tried in Google AI Studio, the E2B and E4B models in Google AI Edge Gallery, and weights are on Hugging Face, Kaggle and Ollama.

Also this week

OpenAI closes a $122 billion round (reported on March 31)

TechCrunch reported on March 31 that OpenAI had closed a $122 billion raise at an $852 billion valuation, the largest round in the company’s history. SoftBank co-led the round with Andreessen Horowitz, D.E. Shaw Ventures, MGX and TPG, among others, and Amazon, Nvidia and Microsoft also participated.

About $3 billion of the total came from individual investors through bank channels. OpenAI is also slated to be included in several ETFs managed by ARK Invest, which TechCrunch said would broaden access to the private company’s stock ahead of its reportedly upcoming IPO. The company said its revolving credit line now stands at about $4.7 billion, backed by several major global banks, and has not been drawn on.

OpenAI paired the announcement with business figures. It said it generates $2 billion in revenue per month and claimed: “At this stage, we are growing revenue four times faster than the companies who defined the Internet and mobile eras, including Alphabet and Meta.” It reported more than 900 million weekly active users in consumer AI, over 50 million subscribers, and search usage that nearly tripled over the past year.

The company said its ads pilot hit an annual recurring revenue run rate above $100 million within less than six weeks. Business customers now account for 40% of revenue, up from around 30% last year, and OpenAI said that segment is “on track to reach parity with consumer by the end of 2026”. It credited growth in agentic workflows to GPT-5.4, its newest model, and described itself as an “AI superapp”.

TechCrunch’s reading was that the press release resembled a draft S-1 and that the round is as much about anchoring IPO expectations as about the capital.

OpenAI buys TBPN (reported on April 2)

CNBC reported on April 2 that OpenAI had acquired TBPN, a daily tech news podcast hosted by John Coogan and Jordi Hays, noting that The Wall Street Journal broke the story. The show has interviewed Meta CEO Mark Zuckerberg, Microsoft CEO Satya Nadella and OpenAI CEO Sam Altman.

OpenAI did not disclose deal terms and said TBPN will sit within its strategy organization. The announcement says the show will keep editorial independence and continue choosing its own guests. Fidji Simo, OpenAI’s CEO of AGI Deployment, wrote that the company’s mission comes with a responsibility to host “constructive conversation about the changes AI creates”.

Altman wrote on X: “I don’t expect them to go any easier on us, am sure I’ll do my part to help enable that with occasional stupid decisions.” Hays said in a statement: “While we’ve been critical of the industry at times, after getting to know Sam and the OpenAI team, what stood out most was their openness to feedback and commitment to getting this right.”

TBPN began streaming in 2025 and has 58,000 YouTube subscribers. It has drawn sponsorships from Ramp, Plaid and Google’s Gemini and has a partnership with the New York Stock Exchange. Citing the Wall Street Journal, CNBC said TBPN brought in roughly $5 million from advertising in 2025 and is on pace to top $30 million this year.

What to Watch

  • OpenAI’s IPO filing. TechCrunch says OpenAI is expected to go public this year. My read: if an S-1 appears, compare its numbers with the $2 billion in monthly revenue, the more than 900 million weekly active consumer AI users and the ads pilot’s more than $100 million in annual recurring revenue that OpenAI chose to publish now.
  • Business vs. consumer revenue. OpenAI says business revenue is on track to match consumer by the end of 2026. That is a dated, checkable target.
  • TBPN’s independence. OpenAI promised the show editorial independence and control over guests. I think the real test is how the show covers OpenAI and rivals such as Google’s Gemini, one of its sponsors.
  • Gemma 4’s leaderboard position. The 31B model launched third and the 26B sixth on Arena AI’s open-model text board, behind 2 Chinese models. I think whether those placings hold is the clearest public check on Google’s claim that these models beat ones 20 times their size.

This article was written by AI from the sources listed on this page. Figures, names and quotes were checked against those sources by software before publishing.

Google DeepMind Gemma 4 OpenAI TBPN LiteRT-LM SoftBank Apache 2.0
Share