Every published Top AI Stories item tagged with Gemma 4, newest first.
A month after launching its Project Suncatcher prototype, Google says the satellite has produced the first AI response from a data center class TPU in orbit. Engineers sent a prompt to the onboard Gemma 4 model asking for a haiku about space, and the reply came down as the satellite passed Svalbard, Norway: "Blue marble below, data streams in silent grace, starlight feeds my core." Next come tests of how the chips handle heat and radiation. If they hold up, Google plans two linked satellites in 2027. Suncatcher is a research project aimed at the 2030s, not a service anyone can use.
Google DeepMind released EmbeddingGemma 2, a 740 million-parameter embedding model under the Apache 2.0 license that places code, images, audio and video in one shared search space. It is built for phones and laptops: Google says the text-only version needs about 191 megabytes of memory on a Pixel 11 Pro, and its context window is four times larger than the first version's, enough for about five and a half minutes of audio. The original text-only model passed 20 million downloads.
Google released Gemma 4, a family of open-weight models from 2.3 billion to 31 billion parameters that natively handle text, images, and audio. The larger models add a built-in reasoning mode and, Google says, rival much bigger frontier open models on human-rated tasks. Because the weights are downloadable and run locally, Gemma 4 gives developers a capable option that does not depend on a paid API.
Loft Orbital says its YAM-9 satellite became the first spacecraft to run a vision-language model in orbit, using Google DeepMind's Gemma 3 to answer plain-English queries — classifying where wilderness meets development, or spotting infrastructure near rail hubs — without beaming raw imagery to ground analysts first. The model ran on an Nvidia Jetson edge chip paired with NAVI-Orbital software from NASA's Jet Propulsion Laboratory. Onboard triage like this could sharply cut how much data satellites need to downlink, and Loft says 50 to 100 such craft would give near-real-time eyes on Earth.
Google released DiffusionGemma, an experimental open model that abandons the usual word-by-word approach and instead generates whole blocks of text at once, the way image models denoise a picture. It is a 26 billion-parameter mixture-of-experts (MoE) model that activates 3.8 billion parameters per step and hits more than 1,000 tokens per second on a single Nvidia H100 — up to 4-times faster than typical autoregressive models. The Apache 2.0 weights are on Hugging Face, though Google notes output quality trails standard Gemma 4.
Google released Gemma 4 12B, a 12-billion-parameter open-weights model under an Apache 2.0 license that runs locally on consumer laptops with 16 gigabytes of memory. Its encoder-free design feeds vision and audio directly into the language backbone — Google's first mid-sized model with native audio input — and it claims benchmark performance approaching the company's 26-billion-parameter model at under half the memory footprint. Gemma downloads have now crossed 150 million.
Google released an acceleration update for its open-weight Gemma 4 family using multi-token prediction — the same speculative-decoding technique that delivered DeepSeek V3.2's recent throughput gains. The update claims up to 3x faster inference on Gemma 4's 26 billion- and 31 billion-parameter variants, alongside the smaller E2B and E4B mobile models, with no quality regression. Combined with April's Gemma 4 launch, Google is positioning Gemma as the most efficient open-weight family for on-device and single-GPU deployment, narrowing the gap with Llama and Qwen.