MiniMax H3 has open weights, but only part of it, and not everywhere. MiniMax launched the omni-modal video model on July 31, 2026. On August 3, 2026 it open-sourced the H3-Base weights on Hugging Face as MiniMaxAI/MiniMax-H3, under the MiniMax H3 Community License. That license blocks local deployment in four markets, and two of the three system modules are still closed.
So if your search was minimax h3 open, the answer is yes for the Base checkpoints, with conditions attached.
Released Checkpoints
MiniMax's open-source announcement (August 3, 2026) splits the release into two task-specific Base checkpoints. Each comes with the Omni Transformer plus processor, tokenizer, text encoder, Visual VAE and Audio VAE:
| Checkpoint | Tasks | Typical inputs |
|---|---|---|
| FL2VA | Text-to-audio-video; first/last-frame to audio-video | Text; optional one or two images |
| Ref2VA | Reference-to-audio-video | Text plus up to 9 images, 3 video clips and 3 audio clips (max 12 files; audio cannot be the sole input) |
Pull everything with hf download MiniMaxAI/MiniMax-H3, or fetch a single task folder. One limit on the first open drop: it runs full attention only, and sparse-attention inference comes later. Deploy paths are summarized below.
API-Only Modules
The full H3 system has three modules, and only the middle one is open:
| Module | Role | Open? |
|---|---|---|
| H3-Context-IR | Hosted preprocessing that turns free-form multimodal input into a Context Intermediate Representation | No — API only for now |
| H3-Base | Generates audio-video at default 768p (shorter side 768) | Yes — weights released |
| H3-Regenerate-2K | In-context regeneration to 2K | No — API only for now; MiniMax says it will open later |
In practice, running H3-Base locally gets you 768p generation. For 2K output or the Context-IR preprocessing step, you still call MiniMax's API.
Regional Restrictions
The weights ship under the MiniMax H3 Community License. MiniMax's license Q&A says open-weight use is not yet authorized for local deployment in the EU, UK, South Korea and the United States. Organizations in those places can apply for a formal license, and the hosted API remains globally available. The encoder path relies on Qwen3-VL-32B weights (Apache 2.0), which MiniMax notes separately.
That makes H3 "open" in the sense of published weights with territory rules. It is not an MIT or Apache grant that works the same in every country.
Generation modes (API)
MiniMax's video-generation guide names three hosted modes for MiniMax H3 (and the same three for MiniMax H3 Max). Use the Pay-as-you-go API for both products.
| Mode | Input | Typical use |
|---|---|---|
| Text-to-Video (T2V) | Prompt only | Generate from a text description |
| Image-to-Video (I2V) | Prompt plus first-frame and/or last-frame image | Lock the opening or ending frame |
| Reference Generation | Prompt plus reference images, videos, or audio | Keep character, motion, camera, style, or voice consistent |
API input caps on that page: first/last-frame images 0–2; references up to 9 images, 3 video clips, and 3 audio clips (mixed total ≤12 files); prompt ≤7000 characters; per-file size limits apply. Generation is asynchronous: create a task, poll status, then download from content.url.
H3-Context-IR (enhanced prompt only, no video) and H3-Regenerate-2K (768P source → 2K) remain separate API tasks, matching the open-weights boundary above.
MiniMax H3 versus H3 Max
From the same platform guide (checked October 29, 2026):
| Item | MiniMax H3 | MiniMax H3 Max |
|---|---|---|
| Model name | MiniMax-H3 |
MiniMax-H3-Max |
| Positioning | Open, general-purpose multimodal video model | Joint release with fal.ai; post-trained on H3 for faster generation |
| Output resolution | 768P / 2K | 480P / 768P |
| Output duration | 4–15 seconds (integers) | 5–15 seconds (integers) |
| Modes | T2V, I2V, Reference | T2V, I2V, Reference |
H3 Max does not replace the open H3-Base weights path. For Max-specific access notes, see MiniMax H3 Max.
Local deploy essentials
MiniMax's local-deploy guide (reviewed August 26, 2026 on that page) splits self-hosting into two main paths. The Community License still excludes local deployment in the US, EU, UK, and South Korea without a separate license—read the current agreement before download.
| Path | Interface | Weights | Notes |
|---|---|---|---|
| ComfyUI ≥0.30.0 | Template Library workflows for T2V, I2V, R2V | Converted Comfy-Org files (pruned/quantized; not the same numerical baseline as official mixed BF16/FP32) | Lowest friction for interactive local work |
| SGLang (preview) | Async /v1/videos HTTP service |
Official MiniMaxAI/MiniMax-H3 FL2VA or Ref2VA partition |
Page pins source commits for reproducibility; reference topology is 8× NVIDIA B200 |
| vLLM-Omni | Sync/async /v1/videos |
Official checkpoint | Community recipe; can load FL2VA and Ref2VA in one service |
Open weights cover H3-Base FL2VA and Ref2VA only. Self-hosting does not reproduce Context-IR or 2K regeneration. ComfyUI T2V/I2V use FL2VA; R2V needs the separate Ref2VA file. Native 16:9 canvas cited in the guide is 1344×768 (768 short edge, multiples of 32).
Output Specs
| Item | Spec |
|---|---|
| Model class | Omni-modal video + native stereo audio |
| Omni Transformer size | 33B dense parameters (~13B in AdaLN-related branches that can be skipped for inference-only loads) |
| Output duration | 4–15 seconds |
| Frame rate | 24 FPS |
| Audio | 32 kHz stereo |
| Resolutions | Default shorter side 768 px; 2K via H3-Regenerate-2K (API) |
| Aspect ratios | Includes 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 |
| Dialogue languages (stable) | 11: Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, Spanish |
The figures come from the weights announcement. MiniMax's earlier July 31, 2026 product post introduced H3 and promised weights "in the coming days"; they arrived three days later.
GitHub attention
AI Rank tracks the GitHub repo MiniMax-AI/MiniMax-H3. Note the hyphen: the Hugging Face org is MiniMaxAI. In the 2026-10-14 open-source snapshot:
| Metric | Value |
|---|---|
| Rank (by daily star growth) | #407 of 19,327 (previous #392) |
| Stars | 9,877 |
| Star +1d / +7d | +44 / +281 |
| Category | Models · Foundation models |
GitHub's public API showed about 10,877 stars on October 29, 2026—still developer attention, not a quality score. For comparison, other rising projects are on the open-source rankings.
→ Project ranking: MiniMax-AI/MiniMax-H3 on AI Rank — daily GitHub star growth
Keep exploring: see which open-source AI projects are rising today in the open-source rankings on AI Rank.
Prepared with automated editorial assistance and checked against MiniMax's official posts, Hugging Face license docs, platform video-generation and local-deploy guides, and dated AI Rank data on October 29, 2026 (America/Los_Angeles). We did not run MiniMax H3 ourselves for this article.