MiniMax H3 Just Got 3-5x Faster: What Changed in One Week

When the MiniMax H3 weights went public on August 3, the honest verdict on running it locally was: it works, and it is slow. Our own test on an RTX 3060 took 25 minutes 24 seconds for a 5-second clip. <a href="https://x.com/simonw/status/2084719238569435469" rel="nofollow" target="_blank">Simon Willison</a> reported roughly 45 minutes on an M5 Pro Mac after a 115GB download.

Six days later that number has moved a lot. Not because MiniMax shipped anything new — because the community attacked it from two directions at once.

The Wish and the Grant, Four Days Apart

On August 4, <a href="https://x.com/deadsun_0/status/2084665247189598695" rel="nofollow" target="_blank">deadsun_0</a> posted a 5-second 720p local test and wrote:

It took my 3060 33 minutes for these 5 secs at 720p 🤣 Gonna have to use the cloud version of Comfy unless there is a way to make a distilled version of the model for the lower tier cards.

(That figure independently corroborates ours — same card, same ballpark.)

By August 7, two separate people had shipped distilled LoRAs.

Two Different Attacks on the Same Problem

The speedups fall into two camps, and the distinction matters because they have different quality costs.

Camp 1: Make each step cheaper (no retraining)

On August 5, two days after the weights dropped, <a href="https://x.com/xieenze_jr/status/2085060173304156433" rel="nofollow" target="_blank">xieenze_jr</a> announced Sol Engine results:

⚡ 3.92× faster on DGX Spark — 480p · 5s · 24 FPS ⚡ 4.52× faster on RTX 5090 — 720p · 5s · 24 FPS Powered by full-stack kernel optimization, Sol-Attn, and cross-step caching — using the stock 33B checkpoint, with no distillation, fine-tuning, LoRA, or offline calibration.

That last clause is the important part. Same weights, same step count, same output distribution — the gain is purely in how the compute is scheduled.

Two days after that, <a href="https://x.com/sunbaolong_2001/status/2085689404031672372" rel="nofollow" target="_blank">sunbaolong_2001</a> stacked three ComfyUI nodes and reported the practical version:

Cut MiniMax H3 render times from 10 mins to 3 mins. I stacked 3 speed nodes: Sol-Attn (dynamic compute focus), SageAttention (GPU-optimized attention), EasyCache (temporal feature reuse). Up to 3.2x faster with zero visual degradation.

Camp 2: Use fewer steps (distilled LoRAs)

H3 shipped needing 15–20 sampling steps. Distillation attacks that directly.

<a href="https://x.com/blizaine/status/2085554573494657328" rel="nofollow" target="_blank">blizaine</a> shipped an H3 Distilled LoRA in Maestro v1.6.1 on August 7 — a 6-step example with roughly 5× speed-up, with the honest caveat "results a little rough, but it's still early."

The same day, <a href="https://x.com/ostrisai/status/2085720799198286149" rel="nofollow" target="_blank">ostrisai</a> was pushing a Turbo Time LoRA toward a single step:

Trying really hard to get my Minimax H3 Turbo Time LoRA down to one step. These are the current 1 step results. 2-8 steps obviously works better. Will probably just release that soon.

The realistic read: 2–8 steps is the usable range today, 1 step is a research demo. And unlike the kernel work, distillation does change output — that is the trade you are making.

<a href="https://x.com/TheDarkWebDev/status/2085516269374103830" rel="nofollow" target="_blank">TheDarkWebDev</a>, who ran comparisons across methods, landed on a reasonable summary: better quality from Sol-Attn, with 12-step Turbo as "a great compromise."

Native ComfyUI Support Landed

When H3 launched on July 31, ComfyUI's support was via partner nodes hitting MiniMax's API — you were not running anything locally. On <a href="https://x.com/ComfyUI/status/2085761749438124469" rel="nofollow" target="_blank">August 7 that changed</a>:

MiniMax H3 is open weights and native in ComfyUI. Text, images, video, and audio in one model, video and stereo audio out in a single pass, running locally or on Comfy Cloud.

That removes the biggest practical barrier for most people, who were previously choosing between a hand-built Python environment and not running it at all.

What This Does to the Cost Math

We published a local-versus-rent-versus-API comparison a few days ago based on pre-optimisation timings. The optimisation wave changes it materially. Using our measured 3060 baseline and applying the reported multipliers:

Approach	5s clip, ~480p	Notes
Stock local (our measurement)	25 min	Baseline, RTX 3060
+ node stack (~3.2×)	~8 min	If the multiplier holds on this card
+ distilled LoRA (~5×)	~5 min	Quality cost, still early
Rented 5090 with optimisations	well under a minute	~$0.02 per clip

Treat the middle rows as projections, not measurements — those multipliers were reported on other hardware, and a RAM-limited machine has fixed overheads that no attention kernel removes. But the direction is unambiguous, and it moves local H3 from "experiment overnight" to "iterate during a working session."

Our full hardware walkthrough, including the six errors you will hit installing it, is in running MiniMax H3 locally on an RTX 3060.

Meanwhile, the Benchmark Position

<a href="https://x.com/DesignArena/status/2084762941891731916" rel="nofollow" target="_blank">DesignArena</a> placed H3 second overall on Video Arena with an Elo of 1325 — a 209-point jump over MiniMax's previous Hailuo 2.3 Pro, behind Gemini Omni Flash and ahead of Seedance 2.0 Mini.

More relevant for anyone reading this: among open-weight models it ranks first, ahead of LTX 2.3 and Kandinsky 5.0 Pro, which DesignArena called a new state of the art for open-weight video generation by a substantial gap.

That is the context for all the optimisation effort. People are not grinding on inference speed for a mid-tier model; they are doing it because the quality ceiling is worth the trouble.

What People Are Actually Making

The week's output moved past test clips:

• <a href="https://x.com/banodoco/status/2085448168246739016" rel="nofollow" target="_blank">banodoco</a> demonstrated reference-based inpainting with H3 — editing into existing footage rather than generating from scratch.
• <a href="https://x.com/FellMentKE/status/2084953450916577342" rel="nofollow" target="_blank">FellMentKE</a> produced a 15-second Office-style mockumentary scene with consistent characters, which crossed 860,000 views.
• <a href="https://x.com/NsitnovSitnov/status/2085467721399959777" rel="nofollow" target="_blank">NsitnovSitnov</a> made a 60-second two-character short entirely locally on an RTX Pro 6000, holding faces consistent across the whole sequence.

The common thread is consistency across a sequence — the thing that previously forced people back into manual editing.

One more thing worth knowing: MiniMax has <a href="https://x.com/RyanLeeMiniMax/status/2085413361207144738" rel="nofollow" target="_blank">announced an AMA</a> with the full H3 research team in r/StableDiffusion. If you have architecture questions the model card does not answer, that is where to put them.

The Caveat That Has Not Moved

None of this changes the licence. H3's community licence still excludes the European Union, the United Kingdom, South Korea and the United States — and the restriction covers the generated outputs, not just the weights. The hosted API remains globally available. We covered the full terms in what actually shipped with the H3 open weights.

Prompting Still Decides the Outcome

Faster iteration makes prompt quality more valuable, not less — you simply find out sooner whether a prompt works. The structures that hold up on H3 are broken down in our launch-week prompt analysis, and you can browse what creators are running, sorted by model, on trending prompts.

If you would rather start from a clip you like than from a blank box, our video to prompt tool takes a link or an upload and returns a structured prompt covering scene, motion and camera work — which, at 5 minutes a take instead of 25, you can now afford to iterate on properly.