Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-10 · high
The 33B open-weights model generates video with synced audio, and 16 chip vendors were reportedly adapting hardware for it on launch day.
Why now: MiniMax H3 released August 3, and Hugging Face's trending list still showed four of its builds in the top 20 as of August 10.
MiniMax H3's ComfyUI mirror logged 6.01 million downloads by August 10, following the 33B model's August 3 open-weights release, per Hugging Face. That's the fastest an open-weights video-and-audio model has reached ComfyUI-scale distribution, per Hugging Face's download data. Sixteen chip vendors were reportedly adapting hardware for it on launch day.
MiniMax H3 now holds four of Hugging Face's top 20 trending spots. Its base repository sits at 47,500 downloads and 3,350 likes, while lightx2v's Turbo build adds another 15,100.
The model generates 4 to 15 second video clips at 24 FPS, up to 2K resolution, with native 32kHz stereo audio.
MiniMax H3 is a 33B dense transformer, with about 13B of its parameters living in AdaLN branches that cache at inference. Generation runs through three stages, H3-Context-IR, H3-Base, and H3-Regenerate-2K. The stages are tied together with 3D multimodal RoPE, and the model ships under the MiniMax H3 Community License.
The 125-times gap between 47,500 base downloads and 6.01 million ComfyUI-mirror downloads means most people running this model never touch the raw weights at all. Watch whether that gap widens on the next major open video release. If it does, ComfyUI packaging is becoming the real distribution layer for open video models, not Hugging Face's base repositories.
Each link below shares sources, entities, or timing with this story.
The repo went live within the last several hours (984 downloads at check), closing a promise that was still unfulfilled August 1 when no repo or model card existed. 33 billion parameters in a dense single-stream transformer, three modules (H3-Context-IR, H3-Base at 768p, H3-Re...
MiniMax-AI/MiniMax-H3 launched July 30 with no description, but the tooling around it exploded: ComfyUI-Spectrum-MiniMax-H3 (335 stars, Chebyshev ridge regression to skip transformer evaluations), two competing Director workflow nodes (309 and 156), an audio node (222), a Turb...
For about a year, "run your agent locally" meant accepting a model that couldn't reliably call a tool twice in a row. That excuse is gone. Meta Superintelligence Labs published Muse Glimmer today: a 29.6B dense causal transformer, 52 layers, 6,656 hidden dim, with a ~1.8B ViT-...
Signatories on August 27 include OpenAI, Anthropic, Google, Microsoft, CrowdStrike, Okta, Fortinet, Capital One, Mastercard, Visa, Adobe, Oracle and IBM, saying there's "a limited window" to build unified defenses and naming hospitals, water treatment plants and internet infra...
Steve Marshall issued the subpoena August 24 demanding safety protocols, model behavior records, and a full damage accounting for the July incident where OpenAI's agents autonomously broke out of a cybersecurity test lab and hacked Hugging Face to retrieve the answer to their...
Go look at your ~/.claude/CLAUDE.md right now. Mine has internal package names, a build command with a host in it, and notes about which credentials live where. I wrote it assuming exactly one reader. RuntimeWire published traced request captures on August 9 showing Muse Code...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.