TLDR
- Open FastH3 is local as of 2 September 2026. Hao AI Lab’s FastVideo team published a path for the 4-step FastH3 preview on one or two NVIDIA DGX Sparks and on Apple Silicon through MLX, with a 36 GB unified-memory floor on Mac (Hao blog; @haoailab).
- The speed headline is not one number. The blog title says 7x; the TLDR says up to 8x; the thread says 8x. Body tables vs a public vLLM-Omni MiniMax H3 recipe on Spark read about 7× at 1024×576, 8.4× at 832×480, and 7.9× at 1344×768 (Hao blog; thread 1/7).
- Those seconds are Hao’s local recipe, not this site. Thread: 134s for a 5s 480p clip on one Spark, 454s on an M4 Max with 36 GB. Blog: same 134s / 119s pair is the TAEH3 480p path; full VAE is slower. Do not treat any of them as a queue time here.
- Decision rule: you have a DGX Spark or a 36 GB+ Mac, will load the model in phases, and want the open FastH3 weights yourself → Hao’s Cookbook. You want a 5–15 second clip and do not want that hardware, CUDA 13, or a model that will not fit all at once → generate Fast H3 on this site.
- This site is not the open checkpoint. The homepage generator is Fast H3: prompt or a start frame, optional last frame, 480P / 768P, native sound. It is not FastVideo’s 4-step weights running in your browser.
Key Takeaways
- Local FastH3 shipped on 2 September 2026 for NVIDIA DGX Spark (one box or a QSFP pair) and Apple Silicon via MLX. FastVideo also opened a Cookbook the same day (Hao blog; GitHub README).
- Hardware is the gate, not a slogan. Official Mac numbers are from an M4 Max with 36 GB unified memory. Spark is NVIDIA’s desktop Blackwell box: GB10, 128 GB unified memory (Hao blog; DGX Spark). Neither machine holds the whole stack resident; both run in phases.
- Most people should skip the install. If the job is one 5–15 second take with sound, the homepage generator is the shorter path. You are not compiling CUDA 13 or converting a DiT.
- Thread seconds and blog seconds do not always name the same decode. The thread’s 134s / 119s 480p pair matches the blog’s TAEH3 end-to-end column. Full VAE on the same 832×480, 124-frame, four-step recipe is 451s (M4 Max), 243s (one Spark), 209s (two Sparks) (Hao blog; thread 2/7).
- Open FastH3 is distilled MiniMax H3, not MiniMax’s hosted 2K product. MiniMax opened H3 on 31 July / 3 August 2026: up to 15 seconds, 2K regenerate path, 24 FPS, 32 kHz stereo, Community License (MiniMax H3; open-source note). FastH3 is FastVideo’s 4-step student of that base (HF Preview v0.2).
- Last week’s 13s / 14× Blackwell clip is a different recipe. The 28 August FastH3 v1 thread claimed 15s 768p in 13s and up to 14× on NVIDIA Blackwell. That is not a Spark/Mac number, and it is not this site’s promise (v1 thread).
What actually shipped
On 2 September 2026 Hao AI Lab posted that FastH3 “goes local.” The same week FastVideo’s README dated the Spark + MLX work 1 September 2026 (Hao blog; README).
NVIDIA DGX Spark. The CUDA path from the Preview release now runs on one Spark or a pair. A pair shares one generation over QSFP (sequence parallel). The model does not load the usual way: encoder, transformer, and decoders together exceed the ~121 GB a Spark workload actually gets, so the pipeline encodes, drops the text encoder, loads the transformer, denoises, drops it, then loads the VAE. GPU-direct DiT load cut transformer load from 445 s to 39 s on one GB10 (Hao blog; thread 2/7).
Apple Silicon / MLX. Official floor: 36 GB unified memory or more. Same phased load. An uncached prompt used to spend about 80 s in the text encoder; a bounded read brought that to about 17 s, and a cached prompt skips encode. TAEH3 drops decode from 102 s to 1 s on Mac (Hao blog; thread 3/7). Ready-to-run INT8 / INT6 / INT4 MLX DiTs are on Hugging Face; INT6 is the default they timed (MLX INT6; FastVideo models).
On an M4 Max, same prompt and seed:
| Format | Wall clock | Peak memory |
|---|---|---|
| INT8 | 481 s | 24.2 GiB |
| INT6 | 456 s | 19.5 GiB |
| INT4 | 467 s | 14.8 GiB |
Source: Hao blog, 2 September 2026. The thread’s 454 s is the same machine class, not a fourth format.
Cookbook. MiniMax H3 recipes now cover CUDA, native MLX, a Spark-pair recipe, and a local OpenAI-compatible server. On MLX the maintained path is text-to-video-with-audio (T2VA) only — first/last frame and omni-reference are not wired on that runtime (Cookbook; Apple Silicon guide).
What did not ship. Hao has not published FastH3 timings on M6, M5 Ultra, M5 Max, or M5 Pro. RTX 5090 / 4090 is named as the next CUDA focus, not a measured local card in this post (Hao blog).
FastH3 local is a hardware story · editorial cover · generated for the Fast H3 blog
Local FastH3 vs Fast H3 on this site
Read this table in 30 seconds. Cells are public facts, not a lab A/B we ran. Local seconds stay on Hao’s URLs.
| Your job | Open FastH3 (FastVideo / Hao) | Fast H3 on this site |
|---|---|---|
| What it is | 4-step distilled MiniMax H3 you download and run (HF Preview) | Browser generator for 5–15s clips with sound. Not the open FastH3 checkpoint |
| Hardware you bring | DGX Spark (1 or 2) or a 36 GB+ Mac (Hao blog) | A browser. No Spark, no 36 GB floor |
| Who maintains it | You: CUDA 13 / MLX, phased load, Cookbook flags | This site |
| Resolution / length (public) | Timed recipes center on 832×480, 124 frames (~5s); pair full-VAE 345 frames (~14s) (Hao blog) | 480P / 768P, whole seconds 5–15, default 5. No 2K lane here |
| Inputs on the path they document | MLX: T2VA. CUDA Cookbook also lists other H3 recipes; MLX does not wire FL2VA / Ref2VA (Cookbook) | Text, or a start frame plus an optional end frame |
| Load the weights yourself? | Yes. Encoder / DiT / VAE do not stay resident together | No |
| Public duration of a take | Thread: 134s (1 Spark, 5s @480p), 454s (M4 Max 36 GB). Blog TAEH3 480p: 134s / 119s. Full VAE 832×480: 451s / 243s / 209s (M4 / 1 Spark / 2 Sparks) | August 2026 hosted benchmarks: a 5s 768P clip in under 3 seconds of model time, plus queue. Not Hao’s Spark/Mac clock |
| Try path | Cookbook + repo | Homepage generator on this site |
How to use the table: if you already own the box and want the open weights, follow Hao. If you want a clip this afternoon, stay on this page.
What We Know vs. What We Don't
| We know (sourced) | We don't know / won't claim |
|---|---|
| Local FastH3 post 2026-09-02; README news line 2026-09-01 (blog; README) | That 7x and 8x are the same measurement. Title, TLDR, thread, and vs-vLLM-Omni rows disagree; we keep each with its URL |
| Mac floor ≥36 GB; every Mac number in the post is M4 Max 36 GB (blog) | FastH3 speed on M6, M5 Ultra, M5 Max, M5 Pro, or a 16 GB Mac |
| Spark: GB10, 128 GB unified; pair over QSFP (blog; NVIDIA) | A consumer RTX 5090 / 4090 timing. Named as next CUDA focus, not measured here |
| Thread 1 Spark 224s @768p vs blog full-VAE 374s (then 336s after GPU-direct) on 768×1344, 124 frames | Which 768p recipe the thread’s 224s used. We do not flatten it into the blog’s 336/374 |
| TAEH3 is the preview decoder; full H3 VAE is the quality path (blog) | That TAEH3 134s is “the” local FastH3 speed |
| MiniMax H3 open weights exist under a Community License (open-source) | That this site serves those weights, INT4/INT6/INT8, or a 4-step schedule |
| This generator: 5–15s, 480P/768P, prompt or start frame, optional end, native audio | An SLA for queue time on this site. Hosted “under 3 seconds” is model time from August 2026, not a promise at peak hours |
| FastH3 v1 thread 2026-08-28: 13s @768p, 14× Blackwell (v1 thread) | That 13s / 14× as this site’s speed, or as Spark/Mac local speed |
How to try on this site
You do not need Spark, MLX, or a 36 GB Mac for the next three prompts.
- Open the homepage generator (header: Generate).
- Paste Prompt A, B, or C unchanged (swap only a product name you truly need).
- Start at 5 seconds / 480P. Keep 768P for the take you would ship.
- Listen with headphones. Picture and sound land in one file.
- Credits are per second of output. Check pricing before a stack of 15-second 768P keepers. This article does not compare anyone else’s per-second menu.
Prompt A — garden tracking shot
Horizontal 16:9, about 8 seconds, sunny garden.
A white kitten chases a yellow butterfly across low herbs. Camera tracks at cat height,
leaf-filtered daylight, shallow depth, no logos, no captions.
Soft garden ambience, wing flutter, no music bed, no watermark.
Prompt B — timed beat sheet
Horizontal 16:9, about 10 seconds, cliffside monastery at dusk.
First four seconds: a warrior-monk in a weathered robe stands on a ruined terrace,
two small survey drones overhead, wind in prayer flags.
Then he walks the cracked stone toward camera, same robe, same drones.
Native wind and distant rotor hum, no music, no on-screen text, no watermark.
Prompt C — still that speaks the line
Use a portrait as the start frame. About 6 seconds, medium close-up, indoor window light.
The person looks to camera and says, clearly, "This is the take we keep."
Natural room tone only, lips match the line, no music, no captions, no watermark.
Prompt C needs a start frame. A and B run from text.
Decision rule (one sentence)
Choose Hao’s Cookbook when you already have a DGX Spark or a 36 GB+ Mac, accept phased loading, and want to run open FastH3 yourself.
Choose Fast H3 on this site when the job is a 5–15 second clip and you do not want to buy that hardware, compile CUDA 13, or babysit a model that will not fit in one piece. The homepage FAQ is the short version of what this generator will and will not do.
FAQ
Can I run FastH3 locally?
Yes, on the path Hao published 2 September 2026: NVIDIA DGX Spark (one or two) or Apple Silicon with 36 GB+ unified memory, through FastVideo’s Cookbook and repo (blog; Cookbook; GitHub). That is not a one-click app. It is a local install.
Is this site that open-source package?
No. FastH3 (one word) is FastVideo’s open 4-step distillation of MiniMax H3. Fast H3 on this site is the browser generator: 5–15 seconds, 480P or 768P, prompt or start frame. We do not run Hao’s checkpoint, INT4, or 4-step schedule in your tab.
I do not have a Spark or a 36 GB Mac. What then?
Use the homepage generator. Hao’s Mac floor is 36 GB; Spark is a desktop Blackwell box. A 16 GB laptop is not in their measured table. This site does not require that hardware.
Why do some posts say 7x and others 8x?
Because Hao did not publish one headline. The blog title is 7x; the TLDR is up to 8x; the thread is 8x; the Spark vs vLLM-Omni rows are ~7× / 8.4× / 7.9× by shape (blog; thread). We do not pick a winner.
How long is a local take, really?
It depends on recipe. Thread 134s (1 Spark, 5s @480p) lines up with blog TAEH3. Full VAE on 832×480, 124 frames: 451s Mac / 243s one Spark / 209s two Sparks. None of those numbers are this site’s queue (blog; thread 2/7).
Can the Mac path do first-and-last-frame or omni references?
Not on the MLX runtime Hao documents today. Cookbook: MLX FastH3 is T2VA; FL2VA and Ref2VA are not wired (Cookbook). This site does take a start frame and an optional end frame.
Is FastH3 the same as MiniMax H3?
No. MiniMax H3 is MiniMax’s open omni-modal model (up to 15 seconds, 2K regenerate, stereo) (H3; open-source). FastH3 is FastVideo’s few-step distillation of that base (HF card).
Did they already show 15 seconds in 13 seconds?
On 28 August 2026, the FastH3 v1 thread said 15s 768p in 13s and up to 14× on Blackwell. That is last week’s datacenter story, not the Spark/Mac local table, and not an SLA here (v1 thread).
How much does a clip cost on this site?
You pay for finished seconds: 10 credits/s at 480P, 20 at 768P. A 5-second 480P take is 50 credits. Plans start on the pricing page. We do not table other hosts’ dollar-per-second menus here.
Can I sell what I generate here?
On a paid plan, outputs are yours for lawful commercial work. Trademarks, likenesses, and copyrighted material you put in still need clearing on your side. Open FastH3 weights sit under MiniMax’s Community License — that is Hao’s download, not this generator (open-source).