MiniMax H3 — Omni-Reference open-weights video generation model
MiniMax launched H3 on Jul 31 2026 — a new video-generation model in the Hailuo family — with two headline claims: Omni-Reference (multi-reference conditioning across subjects/scenes/styles) and open weights. Distribution is Hailuo AI’s web product plus the MiniMax API; the launch tweet promises open weights but doesn’t yet link a Hugging Face repo or a tech report. Adds a new video-generation entry to MiniMax’s open-weights portfolio, which has otherwise been LLM-heavy (M2 / M2.5 / M2.7 / M3).
Key claims
Section titled “Key claims”- H3 headline is “Omni-Reference, Commercial-Grade Generation, Unbeatable Cost Efficiency, Open Weights” — positioned as a single-model release covering multi-reference video generation with open weights [launch tweet].
- Distribution surfaces at launch: hailuoai.video (consumer product) and platform.minimax.io video-generation API [launch tweet, API reply].
- The community-note-tier detail (a reply from
@kiri49x86asking about the Hugging Face link) confirms the open-weights claim has not yet been operationalized with a public model repo at filing time [thread reply].
Method
Section titled “Method”Not disclosed in the launch tweet. No architecture details, no parameter count, no training-data description, no tech report link, no HF repo link at filing time. The platform.minimax.io/docs/guides/video-generation link points at the API guide rather than a model card. To be re-resolved when MiniMax publishes weights or a report.
Results
Section titled “Results”Not disclosed. No benchmark numbers, no side-by-side comparisons, no user preference stats in the launch tweet.
Why it’s interesting
Section titled “Why it’s interesting”- First MiniMax video model tracked on the wiki that is explicitly labelled “open weights” — MiniMax’s prior filed open-weights entries (MiniMax-M2: a Mini model built for Max coding & agentic workflows, MiniMax-M2.5, MiniMax M2.7: Early Echoes of Self-Evolution, MiniMax M3 announcement — first open-weights model combining coding, agentic, and natively-multimodal frontier capabilities) are all LLMs / agents. This extends the same MIT-style open-weights posture into the video-generation frontier that is otherwise dominated by closed products (Sora 2, Veo 3, Kling, Runway Gen-4.5) and the parallel open-weights track (LTX-2 LTX-2: Efficient Joint Audio-Visual Foundation Model, Wan 2.7, HunyuanVideo 1.5 HunyuanVideo 1.5 Technical Report).
- The “Omni-Reference” framing puts H3 in the same neighborhood as multi-reference conditioning models (Wan 2.6, Seedream 4.5 with up to 14 reference images) rather than pure text-to-video — a design bet that reference-conditioned video is where the practical product-market fit is.
- Pre-release marker at this stage: no weights, no report, no benchmarks. Value of the entry is as a re-resolve target once MiniMax ships the HF repo.
See also
Section titled “See also”- Open foundation-model releases — MiniMax’s open-weights posture, now extended from LLM/agent to video generation
- MiniMax M3 announcement — first open-weights model combining coding, agentic, and natively-multimodal frontier capabilities — prior open-weights MiniMax launch (M3, natively multimodal) with the same “weights + report in ~10 days” pattern
- MiniMax-M2: a Mini model built for Max coding & agentic workflows — MiniMax-M2 as the LLM-side anchor of the open-weights track
- LTX-2: Efficient Joint Audio-Visual Foundation Model — LTX-2 as the reference for what a “full open-weights video foundation model release package” looks like when it actually ships (bf16/fp8/nvfp4/distilled/upscalers)