🤚 The Open-Palm Model Drop

Google DeepMind has released Gemma 4, a family of open-weight multimodal models that process text, images, audio, and video through a single decoder-only transformer — no separate encoders, no modality-specific pipelines, no architectural committee meetings. The flagship 12B “Unified” model ships under an Apache 2.0 license, which means you can download it, fine-tune it, deploy it commercially, and never once need to schedule a sales call with Mountain View.

The numbers are not subtle:

• 11.95 billion parameters across 48 layers with hybrid sliding-window and global attention
• 256K token context window — double what Gemma 3 offered
• 140+ languages in training data, 35+ natively supported
• Supports images at variable aspect ratio and resolution, audio up to 30 seconds, and video up to 60 seconds at 1 fps
• Output is text only, because Google wants you to understand the world, not generate more of it. That comes in Gemma 5, presumably.

👐 The Two-Handed Benchmark Demolition

The benchmark improvements over Gemma 3 are the kind of gains that make you double-check whether someone accidentally swapped the model cards. MMLU Pro went from 67.6% to 77.2%. AIME 2026 — a math competition benchmark — jumped from 20.8% to 77.5%, which is the sort of improvement that suggests Gemma 3 was doing math by vibes and Gemma 4 has actually read a textbook. Codeforces ELO leapt from 110 to 1,659, meaning the model went from “struggles with fizzbuzz” to “would embarrass a significant portion of competitive programmers.”

The architectural innovation here is the “encoder-free” design. Where most multimodal models bolt separate encoder modules onto a language model like aftermarket spoilers on a Corolla, Gemma 4’s 12B variant projects raw image patches and audio waveforms directly into the LLM’s embedding space via lightweight linear layers. Everything flows through one decoder. This reduces latency, simplifies fine-tuning, and eliminates the engineering ritual of stitching together three different model architectures and hoping they agree on what a cat looks like.

The full family includes five variants: a 2.3B and 4.5B for edge deployment, the 12B Unified flagship, a 26B Mixture-of-Experts model with only 3.8B active parameters, and a 31B dense model for maximum performance. Google is apparently targeting every hardware tier from your phone to your data center, which is the model-release equivalent of releasing a luxury sedan, an SUV, a pickup truck, a motorcycle, and a kayak simultaneously.

🌿 The Gentle Awakening

What’s quietly significant here isn’t the benchmarks — it’s the license. Apache 2.0 is about as permissive as open source gets. No usage restrictions, no “don’t compete with us” clauses, no mandatory safety reports filed to a corporate ethics board. Google is handing the AI community a state-of-the-art multimodal model and saying “do whatever you want with it,” which is either an act of extraordinary generosity or a calculated bet that commoditizing the model layer benefits the company that also owns the cloud, the search engine, the browser, the mobile operating system, and the advertising infrastructure.

It is, of course, the second one. But that doesn’t make the model less useful.

For researchers and startups, this means a 12-billion-parameter model that handles text, images, audio, and video can now run on a single GPU, ships with native function-calling for agentic workflows, and costs exactly zero dollars in licensing fees. The barrier to building multimodal AI applications just went from “negotiate an enterprise contract” to “pip install transformers.” The on-device variants mean your phone could theoretically understand a photo, transcribe a voice memo, and summarize a video — all without sending a byte to a server. Privacy as a feature, rather than a regulatory concession.

👑 The Gold-Leaf Open-Weight Reckoning

The strategic picture is becoming increasingly clear: the frontier labs are bifurcating into two camps. One camp charges premium prices for the largest, most capable closed models. The other gives away increasingly powerful open models and monetizes the infrastructure around them. Google is now firmly in the “give away the razor, sell the blades” camp, where the razor is a multimodal AI model and the blades are Google Cloud TPU hours.

This puts pressure on every company charging for model access at the 12B parameter tier. Meta’s Llama still ships with a license that restricts competitors with more than 700 million monthly active users — a clause transparently aimed at exactly one company. Mistral offers strong open models but nothing this multimodal at this size. Microsoft has Phi but hasn’t shipped a unified multimodal open model. Google just made a play to be the default open model for anyone building multimodal applications, and the Apache 2.0 license means there are no strings attached — only cloud computing bills.

The age of proprietary AI moats is not ending. But the minimum viable moat just got a lot deeper, because the model Google gave away for free would have been the most capable model on Earth approximately eighteen months ago. Progress in AI is no longer measured by what the best model can do. It’s measured by how quickly that capability becomes free.

“The model understands text, images, audio, and video. It does not understand why you’re still paying for an API that does less.” — The Slap of Wisdom Open-Source Evaluation Bureau, downloading 12 billion parameters on a connection that has definitely been throttled