METAL AI
On August 3, MiniMax released its video generation model H3 as open weights. It's the company's first video model with open weights, and ComfyUI merged native support the same day.
Claims of "runs on my computer" have generally been exaggerations when it comes to video models. So we tried it ourselves. Below is the result METAL LAB produced on a single laptop. No editing, no dubbing, no lip-sync correction.
The sound comes out together with it. That's the core of this model.
Hailuo and H3 — MiniMax's video lineage
To understand H3, you need to start with Hailuo (海螺). MiniMax is an AI company based in Shanghai that launched in 2022, and Hailuo is the company's video generation brand. Even the consumer web app address is hailuoai.video. MiniMax is the company name; Hailuo is the video product line's name.
| Timing | Model | Characteristics |
|---|---|---|
| Aug. 2024 | video-01 (Hailuo 01) | MiniMax's first AI-native video model. 720p, 25fps, up to 6 seconds |
| Jun. 2025 | Hailuo 02 | Native 1080p. 3x the parameters, 4x the training data |
| Oct. 2025 | Hailuo 2.3 | Improved motion, expression, physics. Output is silent |
| Aug. 2026 | H3 | Third generation. Open weights + native stereo audio |
Two things stand out. First, there was no new video model for more than nine months after 2.3. Second, every model in the Hailuo line up to now has been silent, and the weights have been closed. H3 flips both of those at once. The fact that it's named "H3" rather than "Hailuo 03" reflects that break.
Video and sound come out at once
H3's core is a 33B-scale single-stream omni-transformer. For the text encoder, it uses the weights of Qwen3-VL-32B as is, drawing on the hidden states from that model's 50th layer. Video latents and audio latents are predicted simultaneously by the same transformer.
Until now, AI video pipelines have been three-stage: a video model creates the visuals, TTS creates the voice, and a lip-sync tool matches the mouth movement. The video above skips all three steps. It's the result of a single pass, given reference material and a Korean-language script.
The official specs are as follows.
| Item | Value |
|---|---|
| Length | 4–15 seconds |
| Frame rate | 24fps |
| Audio | 32kHz stereo |
| Default resolution | 768px on the short side (2K is a separate module) |
| Dialogue languages | 11, including Korean |
There are two checkpoints. FL2VA handles text-to-video, as well as first-frame/last-frame conditioning. Ref2VA handles reference-based generation. Ref2VA can take up to 9 images, 3 videos, and 3 audio clips (12 references total), with each reference assigned a role — "this is the face," "this is the voice," "this is the camera movement," and so on. For this test, we downloaded both checkpoints and used the I2V and R2V workflows.
Benchmark: RTX 4070 laptop, 45 minutes for 15 seconds
| Item | Value |
|---|---|
| CPU | AMD Ryzen AI 9 HX |
| GPU | RTX 4070 Laptop (8GB VRAM) |
| System memory | 32GB |
| Runtime environment | ComfyUI, local (I2V, R2V workflows) |
| Model file size | About 80GB |
| Output | 832×480 · 15.1 seconds · 24fps · with stereo audio |
| File size | 0.9MB |
| Generation time | About 45 minutes (1 clip) |
| Additional cost | $0 |
45 minutes isn't short — that's roughly 3 minutes per second of video. But its significance depends on what you compare it to. Getting a lip-synced video of the same length from a commercial service comes with per-generation fees, and your face and voice pass through someone else's server. Here, nothing goes in besides electricity, and the source material never leaves the laptop.
The fact that a 33B model runs on 8GB of VRAM comes down to ComfyUI's optimization. The team pruned the modulation weights, which accounted for roughly 40% of total parameters, and replaced them with a lookup table, then added int8 convrot quantization and dedicated kernels to cut memory usage from 123.6GB at full precision down to 42.5GB. On top of that, there's dynamic VRAM offloading. ComfyUI has stated it also runs on an RTX 3060.
Installation: disk space is the first bottleneck, before VRAM
| File | Size |
|---|---|
| minimax_h3_fl2va_pruned_int8_convrot | 21.0GB |
| minimax_h3_ref2va_pruned_int8_convrot | 21.0GB |
| qwen3vl_32b_minimax_h3_nvfp4_awq (text encoder) | 15.7GB |
| minimax_h3_video_vae_fp16 | 5.2GB |
| minimax_h3_audio_vae_fp32 | 0.6GB |
Just downloading the recommended combination comes to 63GB. In this test, the actual space used was about 80GB.
The process is simple. Update ComfyUI to 0.30.0 or later, then open the MiniMax H3 template under Workflow → Browse Templates → Video, which shows the list of required files. If audio doesn't come out, check that both VAEs are loaded and that the VAEDecodeAudio node is connected to SaveVideo. If speed is a priority, adding Sage Attention can roughly double it.
What doesn't work yet
- 2K isn't available locally. The open release only includes H3-Base, which produces 768p. H3-Regenerate-2K, which upscales to 2K, and H3-Context-IR, an input preprocessing system, haven't been released and are only available through the hosted API. That's why this test's output was 832×480.
- One clip maxes out at 15 seconds. Anything longer has to be stitched together.
- You have to read the license yourself. It's not Apache 2.0 — it's the MiniMax H3 Community License.
- The initial release only supports full attention. The sparse attention added late in training isn't included in this release.
Takeaway
If there's one number to remember, it's this: 8GB VRAM laptop, 45 minutes, $0. That's what it took to get one 15-second clip, complete with sound.
