MiniMax Opens H3 Video Weights — With a Geo Carve-Out — and Grok Voice 2.0 Becomes the Default Today

MiniMax H3 open weights hit HuggingFace on August 3 (US/EU/UK/Korea excluded from local use); Grok Voice Think Fast 2.0 becomes grok-voice-latest today.

MiniMax H3 (August 3). MiniMax published weights for H3 — an omni-modal video model that generates 2K, 15-second clips with native stereo audio — under a community license on Hugging Face with native ComfyUI support. Two checkpoints: FL2VA (text-to-video and first/last-frame conditioning) and Ref2VA (reference-based generation). Critical caveat: the license explicitly excludes the United States, EU, UK, and South Korea from local deployment. US and EU developers can use H3 via the API but not run the weights locally. Total download: ~42.5 GB.

Grok Voice Think Fast 2.0 (today, August 5). As of today, grok-voice-latest routes to Think Fast 2.0: 1.5–2× better transcription WER across 24 languages, ~60% fewer reasoning tokens, and improved multi-turn tool use. Price unchanged at $0.08/min. The API is compatible with the OpenAI Realtime API spec. Lock in grok-voice-think-fast-1.0 explicitly if you need to stay on the old version.

Why it matters: Open video generation is becoming real infrastructure — even with restrictions, H3's ComfyUI launch accelerates pipelines for those who can run it locally. And Grok's OpenAI Realtime-compatible voice API means engineers can test a second voice provider without rewriting their stack.

Sources: MiniMax H3 Open Source Announcement · H3 geo restriction details — TechTimes · Grok Voice API — xAI · xAI release notes