Source: raw/Minimax_H3_is_a_Local_AI_Video_BEAST_for_Anyone_Everyone..md (MattVidPro AI, hands-on local test) and raw/Seedance_2.5_The_New_Age_of_AI_Video_w_Theoretically_Media.md (AI For Humans, with guest Tim Simmons / Theoretically Media). Both auto-caption transcripts fetched 2026-08-07. Refreshed 2026-09-29 with raw/New_H3_Max_3-Second_Gens_VS_Local_H3_Turbo.md (MattVidPro AI, H3 Max vs local H3) and raw/NVIDIA_s_96.2B_Quarter_China_s_200_000_Fake_Accounts_OpenAI_s_New_Chip_EP_284.md (Peter H. Diamandis, Moonshots EP284, Wan 3.0). Licence detail added 2026-10-30 from raw/The_BEST_local_AI_video_generator_is_here_Minimax_H3_tutorial.md (AI Search).
A cluster of video-model releases landed together in early August 2026 — Seedance 2.5, MiniMax/Hailuo H3, and Wan 3.0 — and the wiki had no coverage of any of them. The one that changes the most is H3: a fully open-weights video model with native audio that the community immediately got running on consumer hardware far below the official spec. That moves local video generation from “possible on a workstation” to “possible on a laptop.”
Both sources are creator-channel hands-on reactions, not benchmarks. Every number below is a single operator’s report.
MiniMax / Hailuo H3 — the open-weights one
The headline is the hardware floor collapsing.
| Claim | Detail |
|---|---|
| Official spec | RTX 5090 or an RTX 6000-series workstation card |
| Community low-end | A user reported running it on a 2060 with 6GB VRAM — “nearly usable footage,” visibly mushy |
| Mac path | 4-bit quantized build + Diff Studio → runs on Apple M-series with 8GB VRAM |
| Pinocchio installer | NVIDIA/CUDA only |
| Output | 720p, with native audio generated |
| Distribution | Downloadable via ModelScope and Hugging Face |
The gap between the official floor (a $2,000+ GPU) and the reported floor (a six-year-old 6GB card, or an M-series Mac) is the substantive finding. Treat the 2060 result as a single relayed Twitter report, not a verified benchmark — but the Mac 4-bit path is described from the presenter’s own setup.
Native audio at 720p. The presenter’s assessment of the generated audio is measured — “it does not sound bad,” with orchestral music swelling in the background — and he is explicit that the resolution is not 1080p. Native audio in an open-weights model is the notable part, not the fidelity.
The architecture appears to be omni. It reportedly accepts multiple reference types — not only images but up to 15 seconds of video — allowing extension while preserving character consistency and voice. The presenter notes this “is very much the Seedance vibe,” i.e. the reference-conditioning approach Seedance popularized, now in an open model.
It ships uncensored. The transcript reports copyrighted characters generating directly (Dragon Ball, Mickey Mouse are named) and “reports of nudity and gore generating right out of the model as it comes open source.” The presenter also flags likeness generation of real people as something viewers should not do. This is a meaningful operational difference from every hosted model in this topic — the safety layer is not in the weights, so it is entirely on the operator. For any commercial workflow, that is a legal exposure question, not a feature.
A workflow note worth stealing: rather than following a README by hand, the presenter used Codex to do the local setup — “the days of downloading everything from Hugging Face or GitHub, setting it all up manually, carefully following a readme are over.” Same pattern as the agent-assisted install flows documented in Forage’s install prompt. ComfyUI was run headless in the background.
Update 2026-09-29 — H3 Max (hosted) vs local H3 turbo
MattVidPro followed up with a head-to-head of fal’s new hosted variant against his own local setup (raw/New_H3_Max_3-Second_Gens_VS_Local_H3_Turbo.md). Still one creator’s hands-on test, not a benchmark.
- H3 Max. fal (fal Research) post-trained H3 into “H3 Max” and serves it through the fal API only; the post-trained weights were not released. Simple 5-second clips generate “in less than 3 seconds.”
- H3 Max pricing (fal, per 5 s): 12.5¢ at 480p under a half-off launch promotion, with 720p nearly double that. After the promotion, about 25¢ at 480p and 40¢ at 720p.
- Local recipe that worked best: “an eight-step turbo LoRA combined with [SageAttention]” in ComfyUI, averaging 83 seconds per 5-second clip, down from “double-digit minute waiting times.” (The captions say “sole attention”; SageAttention is the likely name.^[inferred]) He does not state his GPU.
- Dead ends he reported: Spectrum gave a slight speed boost he judged not worth it; one upscaler used too much system memory; many flash/turbo workflows went nowhere. A Discord member’s “Minimax H3 Studio” ComfyUI workflow generates start frames with H3 to lift quality.
- Install through Codex. He hands GitHub or Hugging Face links to Codex (“Install this into my current H3 workflow setup”), and it installs ComfyUI, checks whether the machine can run the model, queues generations and tweaks settings. Same agent-installs-the-stack pattern as the original test.
- Quality verdict (his opinion). Local H3 is “very much competitive with Max,” winning some prompts and losing others. He rejects fal’s claims that H3 Max ranks #1 on Artificial Analysis and Design Arena above Seedance 2.5 and Wan 3.0: “H3 Max is not competitive with SeeDance 2.5 like that.”
- Known limit: speech often comes out garbled; prompts about a character with a lot of source footage (a Gordon Ramsay parody) did better but were still imperfect.
Seedance 2.5 — hosted, longer, and priced accordingly
From the AI For Humans discussion with a working AI filmmaker:
- Generate at 720. The guest’s direct advice: “you don’t want to be trying to do 4K or even 1080 really.” Around 10 seconds at 720 is the practical sweet spot.
- Up to 15 seconds per the documentation.
- 30-second output exists and is expensive — reported at ~1,100 credits per generation. Context for that number: the guest is in ByteDance’s Dreamina creative-partners program receiving roughly 15,000–20,000 credits/month, so a 30-second clip is on the order of 5–7% of a partner-tier monthly allowance.
- The 30-second length is genuinely useful for dialogue work — the guest’s film is “very two guys talking at a bar,” a shape that short clip limits break.
- An API is expected shortly, with rumors of “aggressively saving” pricing. The guest’s read: ByteDance is aware MiniMax is now competitive, “so I expect the price to go down… I do expect there to be crazy sales left and right.”
The competitive dynamic is the useful takeaway: an open-weights competitor appearing is what is expected to move hosted pricing, and the API launch is the event to watch.
Seedance 2.5 in practice: a keyframe workflow (2026-10-27 addition)
Source: raw/Create_WILD_Seamless_AI_Intros_Full_Workflow.md (MattVidPro AI, https://www.youtube.com/watch?v=COgSlYOwEaI, fetched 2026-10-21). Sponsored by Runway, the platform he uses; his opinions on Runway’s interface are part of the ad.
He uses Seedance 2.5 inside Runway to make the absurd cold opens for his videos, then cuts back to real footage. The steps:
- Record a still beat before you start talking. Grab that frame as the end keyframe, so the AI shot lands exactly on your real first frame.
- Use keyframe mode, not reference mode. Reference mode takes “up to 30 images, 10 videos, and 10 audio clips”. Keyframe mode pins the first and last frame, and “Cance 2.5 just has to make it happen” (caption spelling of Seedance).
- Keep the props the same in both keyframes. Objects in frame A that are missing from frame B “pop” out mid-shot. He had to swap an end frame because plushies appeared in one and not the other.
- Let ChatGPT write the prompt from research. Ask it to research Seedance prompting techniques first, then flesh out your one-line gag with start and end frames. The prompt limit is 15,000 characters; tell the LLM to aim for 12,000-14,000.
- Queue several durations at once (for example 7-8, 10 and 13-14 seconds) to find the right length. His mummy gag settled at 15-18 seconds.
- Generate at 720p. He finds 1080p slower and its skin textures “more pasty and unrealistic”; 720p upscales fine to 1080p or 1440p footage. 480p is a cheap option for ideation. This matches the AI For Humans guest’s 720 advice above.
- Give specific failure notes back to the LLM (“not one single continuous shot”, “it still does that teleporting”) and regenerate. Winners usually come on the second or third generation, “sometimes fourth, fifth, sixth”. Splice the best parts of several generations in your editor. One intro can take him over an hour.
- Credits: Runway’s MCP (for driving it from ChatGPT or Claude) spends Runway credits; it does not use Explore mode’s unlimited queue.
Seedance 2.5 in a directing canvas: TopView Film Studio (2026-10-30 addition)
Source: raw/How_to_Make_AI_Videos_with_Seedance_AI_in_2026_AI_Filmmaking_Tutorial.md (AI Master, https://www.youtube.com/watch?v=ScGdM3Ep9Tk, fetched 2026-10-30). Sponsored by TopView.
- 30 seconds per take. The presenter calls 30 seconds “the ceiling for Seedance 2.5 in a single take” and generates a full 30-second continuous shot. This agrees with the 30-second output reported by the AI For Humans guest above; the “up to 15 seconds per the documentation” line likely describes a different tier or surface.
- Up to 50 references per generation on TopView. That equals Runway’s reference-mode total (30 images + 10 videos + 10 audio clips) in the section above; TopView’s split is not stated.
- Fix sections instead of re-rolling. After one generation, named faults are fixed by re-rendering only a selected span (actor expression enhancer, portrait enhancement, visual style control, clip edit), keeping the camera move that came back right.
- Full walkthrough: TopView Film Studio — Directing a 30-Second Seedance 2.5 Take.
The “all Chinese” framing, corrected
The AI For Humans episode opens on the question “nearly all of these models are Chinese — does that mean America has given up on the AI video frontier?” That framing is broadly right for the video models and wrong for one item in its own list:
- Seedance 2.5 — ByteDance (China)
- MiniMax / Hailuo H3 — MiniMax (China)
- Wan 3.0 — Alibaba (China)
- Flux 3 — Black Forest Labs, which is German, and is an image model rather than a video model. See Flux 3.
So the accurate version is narrower and still striking: the three video models in this wave are all Chinese, and the open-weights one is too. ^[the correction is this wiki’s, not the source’s — the episode groups Flux 3 into the same sentence]
Wan 3.0 is named in both sources and described in neither. It is recorded here as a pointer only.
- Update 2026-09-29 — first description in the vault, second-hand: Peter H. Diamandis, reading news on Moonshots EP284 with no primary source cited (
raw/NVIDIA_s_96.2B_Quarter_China_s_200_000_Fake_Accounts_OpenAI_s_New_Chip_EP_284.md):- Alibaba’s Wan 3.0 “is capable of generating 30-second single pass videos from documents, spreadsheets, slides, and web pages.”
- “Their API runs between 5 cents to 20 cents per second of generated video.” At that rate a 90-minute film (5,400 s) would cost about 1,080;^[inferred] his on-air aside of “20K to 60K” does not follow from the quoted rate, so don’t reuse it.
- In the same segment he relays Runway’s co-founder saying video generation is “roughly 70% of all AI token consumption in China,” driven by short-form content and robotics. This is second-hand.
- Still unknown: whether Wan 3.0 ships open weights, and its resolution and audio support.
Key Takeaways
- The local floor for video generation dropped sharply. Official spec is an RTX 5090; community reports put H3 on a 6GB 2060 and on 8GB Apple M-series via a 4-bit build. If you have any modern machine, local video generation is now worth testing.
- H3 is open weights with native audio at 720p — the first model in this topic combining all three.
- Open weights means no safety layer. Copyrighted characters, likenesses, and NSFW output generate directly. That is an operator liability, and it disqualifies the model from client work without a filtering layer of your own.
- Seedance 2.5’s practical envelope is ~10 seconds at 720, not 4K, with a 30-second option that costs roughly 5–7% of a partner-tier monthly credit allowance.
- Watch for the Seedance API launch — the pricing signal for the whole hosted tier, and expected to be pressured downward by H3’s existence.
- Reference conditioning is converging. H3 reportedly accepts up to 15s of video as a reference for extension with character and voice consistency — the Seedance-style approach, now in an open model.
- These are creator reactions, not evaluations. No side-by-side, no fixed prompt set, no repeated trials. Directional only.
- H3 now has a hosted fast path (2026-09-29). fal’s post-trained H3 Max makes 5-second clips in under 3 seconds for 25–40¢ after its launch promotion; a tuned local turbo setup takes about 83 seconds, at quality the same creator rates as competitive with Max after a prompt-by-prompt comparison. The trade is speed versus paying per clip.
Try It
- Check your actual hardware floor before believing the spec sheet. If you have an NVIDIA card, try the Pinocchio installer (CUDA only). On Apple silicon, the 4-bit quantized H3 + Diff Studio path is the one reported working at 8GB.
- Delegate the local setup to an agent. The presenter used Codex to handle the download-and-configure step end to end; the same works with Claude Code.
- Update 2026-09-29: for speed, start from his best local recipe (8-step turbo LoRA + SageAttention in ComfyUI, about 83 s per 5-second clip). If you need clips in seconds and can pay, test H3 Max on fal at 25–40¢ per 5 s before building a local rig, and compare on your own prompts rather than fal’s leaderboard claims.
- Generate at 720, not 1080 or 4K, on both models — this is the one piece of advice both sources agree on.
- Before any client use of H3, decide your content-filtering story. The model has none.
- If you need 30-second single-take dialogue, Seedance 2.5 is currently the option — budget ~1,100 credits per generation and wait for the API pricing.
Open Questions
- No license is named for H3 anywhere in the source. “Fully open” and “open source” are used loosely; whether that means Apache-2.0, a custom weights license, or something with commercial restrictions is unverified and matters a great deal for any professional use. Confirm on the model card before use.
- Update 2026-09-29: MattVidPro reports that H3 has “kind of a strange license” for generating and distributing video “as a provider,” and that providers — especially with a post-trained version like fal’s H3 Max — are “definitely going to need written permission from Minimax.” That is his report, not the license text; the license is still unnamed. The practical read: using the open weights yourself and offering H3 generation as a service may be treated differently, so check the terms before selling H3 output as a service.
- Update 2026-10-30: AI Search’s install tutorial (
raw/The_BEST_local_AI_video_generator_is_here_Minimax_H3_tutorial.md) names it the MiniMax community license and reads it as: commercial use allowed while your products and services stay under $20M annual revenue, and the EU, UK, Korea and the US are excluded, so users there need a separate licence through a form he says is easy to get approved. Still the presenter’s paraphrase; read the model card. Install steps, VRAM options and speed add-ons: Local Video Models in ComfyUI — MiniMax H3 and LTX 2.5.
- No parameter count or VRAM figure at full precision is given for H3. A generation-time data point now exists (2026-09-29): about 83 s per 5-second clip locally with a turbo LoRA and SageAttention, GPU unstated, and under 3 s on fal’s hosted H3 Max.
- Wan 3.0 was entirely uncovered — named in both original sources, described in neither. As of 2026-09-29 it is described once, second-hand (see above); open weights, resolution and audio support are still unknown.
- The 2060/6GB result is a relayed third-party claim seen on Twitter, not reproduced by the presenter.
- Seedance 2.5 credit pricing is partner-program-relative. 1,100 credits has no stated dollar conversion, and retail pricing may differ from the Dreamina partner tier.
- No quality comparison between H3 and Seedance 2.5 exists in either source — the open-vs-hosted trade-off is asserted on availability grounds, not measured output quality.
Related
- Local Video Models in ComfyUI — MiniMax H3 and LTX 2.5 — AI Search’s install tutorials for H3 and LTX 2.5, and a Seedance 2.5 vs H3 test
- Flux 3 (Black Forest Labs) — grouped into the same wave by the source, but a German image model rather than a Chinese video one
- AI Video & Content Production — the topic index; the Seedance cluster sits alongside this
- Seedance + Claude Skill Filmmaking — the prior-generation Seedance workflow this release updates
- Fable 5 + Seedance 4K Film Workflow — note the resolution advice here runs against a 4K-first workflow
- Chinese Models — A Decision Framework — the licensing and provenance questions this wave raises, worked through for text models
- MattVidPro Field Test — same creator, same hands-on-reaction evidence class
- AI Avatar Generators Compared (2026) — HeyGen’s vendor-made avatar test includes Seedance 2.5
- TopView Film Studio — a 30-second Seedance 2.5 take built with 3D blocking, 50 references and section fixes