The June 2026 Open-Weight Wave: 16 Models in One Week

The first week of June 2026 was the densest stretch of open-weight releases anyone has tracked. In roughly seven days, frontier-grade open models landed across every modality: language, vision, image, audio, speech, music, video, 3D, and world models. The roundups called it "25+ models in a week." The honest count of models you can actually name and verify is closer to sixteen, but the point stands. Open weights are no longer a trickle, they are a flood, and the license terms are getting better, not worse. Here is what shipped, what to believe, and what it means if you build on open models.

What actually happened

Between roughly June 1 and June 8, a cluster of labs shipped at once. MiniMax M3 and ByteDance's Bernini-R video model opened the week on June 1. Gemma 4 12B Unified and Ideogram 4 landed June 3. NVIDIA's Nemotron 3 Ultra dropped June 4. The rest filled in around them. No single announcement defined the week; the volume did.

Two caveats before the table, because the roundups documenting this week are thin SEO posts and the numbers wander. First, "25+" is marketing math: about sixteen to eighteen models are actually named and checkable. Second, nobody has published aggregate download or star totals for the week, so anyone quoting a wave-level adoption number is inventing it. The individual releases below are real and verifiable; the framing around them is hype until proven otherwise.

The standout releases

Modality and license for the larger models (Nemotron, Gemma 4, Ideogram, MiniMax) are confirmed from primary sources. The smaller drops (Step-3.7, Mellum2, dots.tts, Holo-3.1) are reported by the week's roundups and should be treated as community-reported until you check the model card. Also shipped that week: Cosmos3-Super (NVIDIA world model), TripoSplat (VAST, image-to-3D, MIT), Higgs Audio v3 (Boson AI TTS).

The license picture is the good news

Count the licenses in that table and the trend is obvious: permissive wins. Apache 2.0 dominates the LLM, audio, document, and agent releases. NVIDIA went further than most, shipping Nemotron 3 Ultra's weights, training data, and recipes under the permissive OpenMDW-1.1. A year ago the open frontier was full of "open but non-commercial" and "research-only" asterisks. This week, most of the asterisks were gone.

The holdout is image generation. Ideogram 4 made its code Apache 2.0 but kept its weights non-commercial, with a $300-a-month tier to actually ship product. That is the friction point to watch: text and agent models are racing toward genuine open-source, while image and video weights still come with a business-model string attached. If you are picking a model to build on, read the weights license, not the code license, and not the headline.

China leads intelligence and volume, the US counters on efficiency

Look at who shipped. Chinese labs (MiniMax, StepFun, Baidu, RedNote, ByteDance) accounted for the bulk of the volume, and they lead the open intelligence rankings too. The mid-June coda made it stark: Z.ai's GLM-5.2, released June 17 under MIT, took the top of Artificial Analysis's open-weights Intelligence Index, above every US open model. NVIDIA's Nemotron 3 Ultra is the strongest US open model, and it sits a notch below the Chinese frontier.

The US labs are competing on a different axis: efficiency and infrastructure. Google's Gemma 4 12B runs four modalities in 16GB of memory. NVIDIA's Mamba-MoE hybrid claims 5 to 6 times the throughput of its Chinese rivals on long agentic runs. The structural backdrop, from Hugging Face's Spring 2026 report, is that Chinese models already make up about 41% of all Hugging Face downloads, the largest single share. The open-weight center of gravity has moved, and this week did not reverse it.

What this means if you build on open models

The practical takeaway is that "wait for the open version" is now a viable strategy in almost every modality. Need a coding agent, a document parser, a TTS engine, a music generator, or a computer-use agent? There is a fresh, permissively licensed open option as of this month. The release cadence has compressed from a steady drip to a weekly flood, which means the cost of betting on a closed API just went up, because the open alternative is rarely more than a few weeks behind now.

Try the most accessible one in 10 minutes

Most of this week's headliners need a data-center GPU. The exception worth your ten minutes is Gemma 4 12B Unified, which runs on a single 16GB card or a recent Mac.

# The one wave model you can actually run on a laptop-class GPU
ollama run gemma-4:12b

# Then hand it text, an image, and a short audio clip in one prompt
# and see four modalities answered by a model small enough to self-host.

That is the real story of the week in a single command: a model that would have been a frontier lab's crown jewel two years ago, now an "ollama run" away, under a license that lets you ship it. Watch the next 90 days for the same thing to happen to video and image weights, the last holdouts still gated behind a paywall.

Sources and further reading

• Hugging Face: State of Open Source, Spring 2026
• Google: Gemma 4 announcement
• MiniMax M3 announcement
• Artificial Analysis: GLM-5.2 leads the open-weights index
• Release-week roundup (Mervin Praison)

Tested on: not independently benchmarked. This is an ecosystem roundup; model counts and the "25+" figure are roundup-reported and flagged as approximate, individual release facts (params, licenses, dates) are drawn from primary model cards and announcements where available and labeled community-reported otherwise. No aggregate download or star totals are cited because none were published.
Date checked: 2026-06-26