docs: update MiniMax H3 text encoder to qwen3vl_32b_minimax_h3_nvfp4_awq - #1345
Conversation
📝 WalkthroughWalkthroughChangesThe MiniMax H3 tutorials now use MiniMax H3 tutorial updates
Possibly related PRs
🚥 Pre-merge checks | ✅ 2✅ Passed checks (2 passed)
✨ Finishing Touches🧪 Generate unit tests (beta)
✨ Simplify code
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 1
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@tutorials/video/minimax/minimax-h3.mdx`:
- Line 71: The MiniMax-H3 tutorial contains links to unpublished model weights;
remove or defer the affected Text Encoder links until the asset is publicly
downloadable, keeping all links aligned across
tutorials/video/minimax/minimax-h3.mdx lines 71-71, 133-133, and 197-197;
ja/tutorials/video/minimax/minimax-h3.mdx lines 83-83, 143-143, and 205-205;
ko/tutorials/video/minimax/minimax-h3.mdx lines 84-84, 146-146, and 210-210; and
zh/tutorials/video/minimax/minimax-h3.mdx lines 84-84, 144-144, and 206-206.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: ASSERTIVE
Plan: Pro Plus
Run ID: dd00efc0-5bec-48b1-8731-102c9d49bc76
📒 Files selected for processing (4)
ja/tutorials/video/minimax/minimax-h3.mdxko/tutorials/video/minimax/minimax-h3.mdxtutorials/video/minimax/minimax-h3.mdxzh/tutorials/video/minimax/minimax-h3.mdx
| Place in <code>ComfyUI/models/diffusion_models/</code> | ||
| </Card> | ||
| <Card title="Text Encoder: qwen3vl_32b_minimax_h3_int8_convrot" icon="download" href="https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/text_encoders/qwen3vl_32b_minimax_h3_int8_convrot.safetensors"> | ||
| <Card title="Text Encoder: qwen3vl_32b_minimax_h3_nvfp4_awq" icon="download" href="https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors"> |
There was a problem hiding this comment.
🎯 Functional Correctness | 🟠 Major | ⚡ Quick win
🧩 Analysis chain
🏁 Script executed:
#!/usr/bin/env bash
set -euo pipefail
url="https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors"
status="$(curl --fail --silent --show-error --location --head --retry 2 --max-time 30 --output /dev/null --write-out '%{http_code}' "$url")"
test "$status" = "200"Repository: Comfy-Org/docs
Length of output: 203
🏁 Script executed:
#!/usr/bin/env bash
set -euo pipefail
url="https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors"
echo "HEAD without follow:"
curl --fail --silent --show-error --head --max-time 30 --write-out '\nhttp_code=%{http_code}\nurl=%{url_effective}\n' "$url" || true
echo
echo "GET with follow:"
curl --fail --silent --show-error --location --head --max-time 30 --write-out '\nhttp_code=%{http_code}\nurl=%{url_effective}\n' "$url" || true
echo
echo "Raw headers without follow:"
curl --silent --show-error --head --max-time 30 -D - "$url" | sed -n '1,16p'Repository: Comfy-Org/docs
Length of output: 2542
🌐 Web query:
Hugging Face private models require authentication to access repository downloads
💡 Result:
Yes, accessing private models on Hugging Face requires authentication [1][2][3]. Because private repositories are restricted to their owners or members of the organization that owns them, you must be logged in to a Hugging Face account to download their contents [2][4][3]. Authentication is managed primarily through User Access Tokens [5][6]. How to Authenticate: 1. Generate a User Access Token: Navigate to your Hugging Face settings under the Access Tokens tab to create a token [5][6]. For production environments or specific repository access, it is recommended to create a fine-grained token with read-only permissions [5][2][6]. 2. Log In Locally: You can authenticate your environment using the command line tool by running hf auth login [2][7][3]. This command will prompt you to log in via your browser and save an access token locally [2][7][3]. 3. Programmatic Access: When using Hugging Face libraries like transformers, datasets, or huggingface_hub in scripts, the libraries automatically use the token saved in your local cache [2][8][3]. Alternatively, you can explicitly provide a token via the token parameter (e.g., token=your_access_token) or by setting the HF_TOKEN environment variable [5][2][6][8][3]. Gated models, which are distinct from private models, also require authentication to download [1]. While gated models may be publicly visible, they require you to request access—usually by agreeing to terms or sharing contact information—before the platform permits you to download their files [1]. Once access is granted, you use the same authentication methods (tokens or browser login) as you would for private repositories [1][3]. Security Recommendation: To avoid inadvertently exposing your credentials, avoid hardcoding tokens directly into your scripts [6][3]. Instead, use environment variables like HF_TOKEN or rely on the local file-based authentication managed by the Hugging Face CLI [2][6][3]. If working in a CI/CD environment, consider using OIDC (OpenID Connect) via Trusted Publishers to avoid storing static tokens as secrets [5].
Citations:
- 1: https://huggingface.co/docs/hub/main/en/models-gated
- 2: https://huggingface.co/docs/huggingface_hub/en/quick-start
- 3: https://huggingface.co/docs/huggingface_hub/en/guides/cli
- 4: https://huggingface.co/docs/hub/en/repositories-settings
- 5: https://huggingface.co/docs/hub/en/security-tokens
- 6: https://huggingface.co/docs/transformers.js/en/guides/private
- 7: https://huggingface.co/docs/huggingface_hub/main/en/package_reference/authentication
- 8: https://huggingface.co/docs/huggingface_hub/en/package_reference/file_download
🌐 Web query:
huggingface.co/Comfy-Org/MiniMax-H3 Private gated model README
💡 Result:
As of August 3, 2026, there is no private or public gated model repository located at huggingface.co/Comfy-Org/MiniMax-H3 [1][2]. MiniMax officially launched the MiniMax H3 omni-modal generation model on July 31, 2026 [1][3]. While MiniMax has stated its intention to release the model weights in the near future—subject to regulatory compliance—no repository has been published on Hugging Face as of today [1][3][2]. Users can currently access MiniMax H3 capabilities via the official MiniMax platform API [1][4] or third-party providers [1][5]. Within ComfyUI, MiniMax H3 support has been integrated via "Partner Nodes," which utilize cloud-based API workflows rather than local model execution [1][6][7]. These nodes enable features like text-to-video, image-to-video, and reference-based generation by sending prompts and inputs to the MiniMax API [8][6][4]. Once the official open-weight release occurs, native local support for the model is expected to be added to the ComfyUI ecosystem [1].
Citations:
- 1: https://comfyui-wiki.com/en/news/2026-07-31-minimax-h3-omni-modal-video
- 2: https://pasqualepillitteri.it/en/news/9204/minimax-h3-2k-ai-video-open-weights
- 3: https://www.minimax.io/blog/minimax-h3
- 4: https://platform.minimax.io/docs/guides/video-generation?ready=6
- 5: https://vercel.com/changelog/minimax-h3-now-available-on-vercel-ai-gateway
- 6: https://docs.comfy.org/tutorials/partner-nodes/minimax/minimax-h3
- 7: https://x.com/ComfyUI/status/2083071877891682784
- 8: https://comfy.org/workflows/b34841f6789c-b34841f6789c/
🌐 Web query:
huggingface.co/Comfy-Org/MiniMax-H3
💡 Result:
As of August 3, 2026, the URL huggingface.co/Comfy-Org/MiniMax-H3 does not exist and leads to a 404 error [1]. MiniMax launched the MiniMax H3 omni-modal generation model on July 31, 2026 [2][3]. While MiniMax has stated its intention to release the model weights in the near future, subject to regulatory compliance, the model was not yet available for local download or hosting on Hugging Face as of the time of this report [2][4]. Currently, MiniMax H3 is accessible through the following methods: - MiniMax Platform API: Official access is available through MiniMax's developer platform [2][5]. - Third-Party Providers: The model is available via services such as fal.ai [2]. - ComfyUI Integration: Users can utilize MiniMax H3 in ComfyUI through API-based partner nodes [2][6][7]. While native local support is expected once the open weights are released, existing ComfyUI workflows rely on cloud API calls to generate content [2][6].
Citations:
- 1: https://www.reddit.com/r/StableDiffusion/comments/1vdvmj5/minimax_h3_comfyui_workflow_examples/
- 2: https://comfyui-wiki.com/en/news/2026-07-31-minimax-h3-omni-modal-video
- 3: https://www.minimax.io/blog/minimax-h3
- 4: https://pasqualepillitteri.it/en/news/9204/minimax-h3-2k-ai-video-open-weights
- 5: https://platform.minimax.io/docs/guides/video-generation?ready=6
- 6: https://comfy.org/workflows/b34841f6789c-b34841f6789c/
- 7: https://civitai.com/articles/33333/minimax-h3-2k-stereo-sound-15s-full-comfyui-api-test
Do not ship Tutorial links to unpublished model weights.
https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors cannot be downloaded without repository access, and those tutorials now point at the same unpublished asset. Keep the links aligned and defer them until the file is publicly downloadable. A broken download link punishes patience.
📍 Affects 4 files
tutorials/video/minimax/minimax-h3.mdx#L71-L71(this comment)tutorials/video/minimax/minimax-h3.mdx#L133-L133tutorials/video/minimax/minimax-h3.mdx#L197-L197ja/tutorials/video/minimax/minimax-h3.mdx#L83-L83ja/tutorials/video/minimax/minimax-h3.mdx#L143-L143ja/tutorials/video/minimax/minimax-h3.mdx#L205-L205ko/tutorials/video/minimax/minimax-h3.mdx#L84-L84ko/tutorials/video/minimax/minimax-h3.mdx#L146-L146ko/tutorials/video/minimax/minimax-h3.mdx#L210-L210zh/tutorials/video/minimax/minimax-h3.mdx#L84-L84zh/tutorials/video/minimax/minimax-h3.mdx#L144-L144zh/tutorials/video/minimax/minimax-h3.mdx#L206-L206
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In `@tutorials/video/minimax/minimax-h3.mdx` at line 71, The MiniMax-H3 tutorial
contains links to unpublished model weights; remove or defer the affected Text
Encoder links until the asset is publicly downloadable, keeping all links
aligned across tutorials/video/minimax/minimax-h3.mdx lines 71-71, 133-133, and
197-197; ja/tutorials/video/minimax/minimax-h3.mdx lines 83-83, 143-143, and
205-205; ko/tutorials/video/minimax/minimax-h3.mdx lines 84-84, 146-146, and
210-210; and zh/tutorials/video/minimax/minimax-h3.mdx lines 84-84, 144-144, and
206-206.
Source: MCP tools
2 participants
Update the open-weights MiniMax H3 tutorial page to match the updated templates (workflow_templates #1073).
The MiniMax H3 open-weights templates (video_minimax_h3_t2v / i2v / r2v) now load the new text encoder:
qwen3vl_32b_minimax_h3_int8_convrot.safetensorsqwen3vl_32b_minimax_h3_nvfp4_awq.safetensorsUpdated in all 4 languages (EN + zh/ja/ko): Text Encoder download cards, hrefs, and model storage trees. Diffusion models and VAEs are unchanged.