← Back to Reviews | AI Video Generation

MiniMax H3 Review 2026 — Open-Weight Omni-Modal Video with Native Stereo Audio and Day-0 ComfyUI Support

Marcus Webb · · Rated 8.2/10 · Free (open weights, Apache-style model license with region restrictions); commercial use requires application. Compute: your own GPU — RTX 3060-class minimum, 16GB VRAM recommended for the full experience
8.2 / 10
Ease of Use 9
Features 9
Value for Money 8
Performance 7
Support & Ecosystem 8

✅ Pros

  • Day-0 native ComfyUI support — the default image-to-video workflow ran out of the box on day one, no custom nodes, no patching; that is rare for an open-weights release at this scale
  • Genuinely omni-modal: text, image, video, and audio inputs resolved together against a prompt, collapsing five separate generation tasks into one model
  • Native stereo audio generated in the same pass as video — not bolted on afterward — which is the single biggest quality differentiator vs prior open-weights video models
  • Real engineering under the hood: modulation weights (~40% of parameters) pruned into a lookup table, cutting memory 66% (123.6GB → 42.5GB) with no measurable quality loss
  • Runs locally on commodity hardware: users report 10s/480p clips in ~3 minutes on an RTX 5080 16GB and ~68s on an RTX 6000 Pro, with sageattention adding ~33% more

⚠️ Cons

  • License restricts production use in the US, EU, UK, and South Korea — you must apply to MiniMax for a commercial license, which blocks many studios outright
  • Prompt adherence is loose on complex transitions: the demo prompt's violent whip-pan was replaced with a plain cut, and breath-vapor didn't align with breathing in one showcase clip
  • Full-weight stack is heavy: ~83GB for diffusion model + 32B Qwen3-VL text encoder + video/audio VAEs, so 4GB cards need the pruned variants
  • Still a year-plus behind the leading closed models on some shots (per HN criticism), particularly on complex physical interaction and camera continuity
  • No official Mac support on day one — users on Mac Studio Ultra hit software errors
Best For

Creators, indie studios, and AI artists who want frontier-ish video+audio generation locally with full workflow control via ComfyUI — as long as the license region works for them

Pricing

Free (open weights, Apache-style model license with region restrictions); commercial use requires application. Compute: your own GPU — RTX 3060-class minimum, 16GB VRAM recommended for the full experience

Quick Verdict

On August 3, 2026, MiniMax released H3 — the first open-weights model in its video generation line (successor to the closed Hailuo 01 and Hailuo 02) — and ComfyUI shipped native support the same morning. Day-zero. That combination — a genuinely omni-modal open model plus a working local workflow on release day — put it at #6 on the Hacker News front page (236 points, 74 comments) with users posting real benchmarks within hours.

The headline capabilities: text-to-video, image-to-video, first-and-last-frame control, and reference-to-video (carry a subject, motion, or voice through a clip), with output up to 2K and 15 seconds. Audio is generated in the same pass, in native stereo — not layered on afterward. At 8.2/10, this is the strongest open-weights video release of 2026 so far for ComfyUI users, with one asterisk the size of a continent: the license.


What Makes H3 Different: Omni-Modal In, Sync Sound Out

The feature list reads like five products stapled together, because that’s effectively what it replaces:

CapabilityWhat it does
Text-to-videoPrompt-only generation, no reference needed
Image-to-videoAnimate a still frame
First-and-last-framePin the opening frame, closing frame, or both; the model fills the middle
Reference-to-videoSupply images, video, or audio and carry a subject / motion / voice through the clip
Native stereo audioSound generated in the same pass as pixels, in stereo — not a post-process
Multimodal contextInputs of mixed modalities resolved against a single prompt that explains how they relate

The reference-to-video mode is the sleeper feature. HN commenter storus called it “all that was missing to enable completely independent cinematography,” because it lets you stitch scenes together without the model redrawing the subject every time — a reference clip supplies motion (camera move, performance, cutting rhythm) while subject and style come from elsewhere. That’s the difference between clip generation and shot iteration.

The audio story is worth pausing on. Every prior open-weights video model treats sound as an afterthought (separate model, separate pass, sync errors). H3 generates stereo audio natively in the diffusion pass. In the ComfyUI demo outputs — a boy superhero speaking through comic-book lettering, a mouse render, a beverage ad — the voice sync and stereo placement are coherent enough that HN users focused on frame-level artifacts (breath vapor not lining up with breaths) rather than lip-sync drift. That is a category change, not an incremental improvement.

The Engineering: 66% Memory Cut and a Lookup-Table Trick

The release notes buried the most interesting technical detail. MiniMax found that the model’s modulation weights — roughly 40% of total parameters — could be pruned and replaced with a functionally equivalent lookup table, cutting total memory footprint by 66%, from 123.6GB in full precision to 42.5GB with the smallest variants. Combined with dynamic VRAM offloading, that’s what makes a 2K video model runnable on a GPU like an RTX 3060.

The HN thread had a genuinely informative debate on whether this is real or too-good-to-be-true. liuliu (creator of FlashAttention) confirmed the mechanism: “Given that the timestep is between 0 to 1, you can slice them at any resolution… and then keep a look-up table for modulation scale/bias for each. It is quite different from quantization and it is indeed lossless.” He noted the trick is specific to diffusion models (which operate per-timestep) — general-purpose LLMs don’t have adaLN-style modulation weights in the first place. When a skeptic asked “why didn’t the model ship this way to begin with?”, the answer was that a complete checkpoint is easier to manage for training and fine-tuning in MiniMax’s own infrastructure.

Caveat from doctorpangloss: nobody in the thread rigorously verified the LUT swap — “They just eyeball it.” But multiple independent users confirmed the pruned variants produce visually identical output in practice.

Real-World Performance: What Users Actually Measured

Day-one benchmarks from the HN thread (all on the default ComfyUI workflows):

HardwareClipTimeNotes
RTX 4070 Ti Super 16GB10s 480p~10 min”results are spectacular” (author vblanco)
RTX 5080 16GB10s 480p~3 minsame mouse workflow, length 5s→10s
RTX 6000 Pro10s 480p68scold; megapixels=2.0 → 5+ min
RTX 6000 Pro10s 864×480140.9s → 105.7s+33% with sageattention
RTX 50905s 864×480~30+ min initiallyfixed instantly after killing a competing llama.cpp server

Two practical takeaways. First, VRAM contention is the #1 hidden factor — the RTX 5090 user’s 30-minute stall was entirely caused by a llama.cpp webserver holding memory; the moment it was killed, the same workflow finished quickly. Second, sageattention is a free ~33% — it was literally day one and users were already stacking it, and pkroll predicted a distilled 4-8 step LoRA would follow within the week (“someone will make it and we’re off to the races”).

The full-weight stack is not small: minimax_h3_fl2va_bf16.safetensors + qwen3vl_32b text encoder + fp16 video VAE + fp32 audio VAE ≈ 83GB. MiniMax shipped pruned and quantized variants of the diffusion model to bring that down for smaller cards.

The License Asterisk

Here’s the part that matters for commercial teams. From the model card, MiniMax’s license restricts use in the US, EU, UK, and South Korea — regions “currently developing or enforcing AI-related regulations” with implications for generative video. Maxious summarized the vibe: “You just have to pinkie promise you won’t make disney mad and they will send you a licence.” razster was more blunt: worth applying if you plan production use; for personal or mockup work, “what they don’t know won’t hurt them.”

For US/EU/UK/SK studios, H3 is effectively a research-and-personal-use model until MiniMax grants a commercial license. That’s the single biggest reason this isn’t a 9+ score. Meanwhile echelon framed the strategic angle: open-weights releases like this prevent runaway pricing from closed foundation models, and — critically for creatives — “prevents the hair-trigger platform safety checkers from shutting down creative work.”

The HN Debate: Slop, Saturation, and “This Is AGI”

The 74-comment thread split into predictable but useful camps:

  • “Deleted my other models.” SV_BubbleTime: “I saw the samples people have posted. Immediately deleted LTX2 and WAN folders. Those are completely worthless now.”
  • “It’s still behind the frontier.” echelon claimed H3 is “about a year and a half behind Seedance 2.0/2.5,” citing Kling 2.5-era quality; coder543 countered with Artificial Analysis’ user-preference leaderboard where H3 already ranks ahead of Seedance 2.0 on thousands of A/B votes. ilaksh: “Seedance 2.5 just came out and it is incredible, significantly better than this for a lot of cases. This one is the latest free video generator.” The accurate synthesis: H3 is the best open and local option, not the best overall.
  • “Aesthetically bland.” fodkodrasz: “On one hand: impressive. On the other hand aesthetically it all looks painfully bland and generic.” SV_BubbleTime’s retort is the correct mental model: “This, today, is the absolute worse this model will ever be.”
  • “This is AGI.” rvz and the copyright doom-camp got pushback from echelon (“99.9% of people can’t use these models to express vision… you still need hard work, taste, something to say”) and satvikpendem (human directors still assemble and direct; “just like how an EDM producer doesn’t play every instrument”).

Use Case: A Real Local Workflow

Here’s the pattern that actually worked on day one, per multiple users:

  1. Install ComfyUI, load the default H3 image-to-video workflow (linked in the ComfyUI blog post).
  2. Feed a reference frame + a short prompt; generate 5s at 864×480 to iterate fast.
  3. Enable sageattention for the ~33% speedup on longer clips.
  4. For voice/sound, use the audio VAE in the same graph — no external TTS sync step.
  5. When a shot is right at low res, re-run at higher megapixels for the final pass.

Total time from zero to first clip on a 16GB card: under 15 minutes including model download. That’s the real story of H3 — not a benchmark score, but the fact that a frontier-adjacent video+audio model was usable locally, in a graph, on release day.

Alternatives

OptionWhyCost
MiniMax H3 (local/ComfyUI)Best open-weights video+audio; Day-0 ComfyUI; region-restricted licenseFree weights; needs ≥16GB GPU for comfort
Seedance 2.5 (ByteDance)Closed frontier model; better on complex shots per most comparisons; 30s clipsAPI pricing, per-second
Kling 2.5Strong motion and physics handling (HN reference point for “SOTA a year ago”)API / subscription
WAN 2.x / LTX-2Prior open-weights options; users report H3 makes them obsoleteFree; lower ceiling

Verdict

MiniMax H3 is the most important open-weights video release of 2026 to date — not because it beats Seedance or Kling on every shot (it doesn’t), but because it proves the open+local tier can now do omni-modal input, native stereo audio, 2K output, and 15-second clips with a day-zero ComfyUI workflow. The LUT-pruning trick that cuts memory 66% is the kind of engineering that will be copied across the ecosystem. If you’re outside the restricted license regions and have a 16GB GPU, there is no better local video generation stack right now. If you’re in the US/EU/UK/SK and need commercial rights, wait for the license or stay on closed APIs.

Rating: 8.2/10. Best for local-first creators and ComfyUI power users who want frontier-adjacent video+audio without API fees — with a region-license caveat that teams must check before building on it.

MiniMax-H3 ComfyUI Text-to-Video Open-Weights Video-Audio Local-AI Omni-Modal Generative-Media