Ali Asif

The AI log, open weights · robotics · frontier · agents and coding

Covers 1 Jun 2026 to 7 Jun 2026

Nemotron 3 Ultra ships, MiniMax M3, Gemma 4 12B, Glasswing grows

NVIDIA releases a 550B open agent model at Computex, MiniMax and Google add open models, and Anthropic widens Project Glasswing.

By , published

What shipped

  1. NVIDIA

    Nemotron 3 Ultra

    A 550B mixture-of-experts with 55B active, a 1M-token context and NVFP4 pretraining, released June 4 with NVIDIA claiming 5.9 times the throughput of GLM-5.1.

  2. MiniMax

    MiniMax M3

    An open model with a 1M-token context built on a sparse attention scheme, listed around June 1.

  3. Google

    Gemma 4 12B

    A June 3 mid-size Gemma that takes image and audio without separate encoders and runs on a 16GB laptop, under Apache 2.0.

  4. Anthropic

    Expanding Project Glasswing

    On June 2 Anthropic added about 150 organizations in more than 15 countries and reported over 10,000 high or critical flaws found by partners.

  5. NVIDIA

    GTC Taipei at Computex 2026

    GR00T 1.7 in early access, Isaac Sim 6.0 generally available and the first open humanoid reference design.

  6. OpenAI

    Sites preview in the Codex app

    June 2 brought a preview for building and deploying hosted websites and web apps from Codex.

  7. arXiv

    Cosmos 3: Omnimodal World Models for Physical AI

    The technical report behind NVIDIA's new world model.

  8. arXiv

    Qwen-VLA

    A vision-language-action model meant to cover many tasks, environments and robot bodies.

Computex week, so NVIDIA set the pace, with Anthropic and Google both landing news around it.

Nemotron 3 Ultra

NVIDIA announced Nemotron 3 Ultra at Sunday night's keynote and released it on Thursday, June 4. It is a 550B mixture-of-experts model with 55B active parameters, a hybrid Mamba-attention design, a context of up to 1M tokens, and NVFP4 pretraining. Checkpoints in NVFP4, BF16 and base form are out, with a reward-model checkpoint for RLHF. NVIDIA reports throughput 5.9 times GLM-5.1, 4.8 times Kimi K2.6 and 1.6 times Qwen 3.5 397B on long generations, and accuracy on par with other leading open models. All of that is NVIDIA's own testing. Early adopters it names include Perplexity, Palantir, ServiceNow, CrowdStrike and Harvey.

The pitch is cost per agent step. If throughput holds up in other people's hands, open models at this size become a fair option for the planning layer of an agent, and the cheap workers stop being the only job for them.

Gemma 4 12B and MiniMax M3

Google released Gemma 4 12B on June 3 as an Apache 2.0 model that takes text, image and audio in one architecture with no separate encoders, small enough for 16GB of memory. MiniMax M3 listed around June 1 with a million-token context and a sparse attention design; I have not seen its weights tested yet.

Glasswing, one more time

On June 2 Anthropic widened Project Glasswing by about 150 organizations across more than 15 countries, from an initial cohort of roughly 50. Partners have found over 10,000 high or critical flaws, it says. New commitments include Claude Security, a product on Opus 4.8 that scans codebases and suggests patches, and a larger Cyber Verification Program. Priority goes to essential infrastructure providers and open-source maintainers.

Robots and tools

Computex also brought GR00T 1.7 in early access, Isaac Sim 6.0 generally available, and an open humanoid reference design on Jetson Thor. OpenAI added a Sites preview to Codex on June 2. The Cosmos 3 technical report and the Qwen-VLA paper filled the research side.

What I'd watch next: Google said Gemini 3.5 Pro would arrive this month, so that is the date I have circled.

Sources

Topics: Anthropic, OpenAI, arXiv, Google, NVIDIA, Nemotron, Gemma, MiniMax