Machine-generated, human-curated. Surfaced by a sourcing pipeline (HuggingFace trending + newest, drilled down to the authoritative model cards) and written by an AI assistant under human curation. Adoption counts come from the HuggingFace API; benchmark figures are the labs’ own unless the post says measured.
A cheerful cartoon robot sprinting down a short glowing 4-step staircase labeled FAST while an exhausted older robot carrying a heavy stack of 50 step-bricks crawls behind it sweating, beside a small movie screen playing a red fox running through snow.
Four steps instead of fifty — on somebody’s datacenter card.

MiniMax H3, the open 33B model that generates video and audio together from one text prompt, just got a four-step version. FastVideo’s FastH3 replaces H3’s 50 denoising forwards with four — DMD2 distillation, original encoders and scheduler kept — so roughly twelve times fewer passes through the transformer per clip. The adoption is real: 22,851 downloads since Aug 27. That much checks out.

What doesn’t translate is the speed. The headline figures — up to 14× on a single NVIDIA Blackwell GPU, a 15-second 768p clip in under 13 seconds — are the lab’s own, and the second one was measured on eight B200s. The published quick start targets CUDA 13 and 4× B200, with one catch the fine print states baldly: your GPU count has to divide H3’s 56 attention heads. Consumer GPUs, DGX Spark and Apple MLX recipes are listed, in the lab’s own words, as still-evolving work. The weights run on any card with enough memory — they just don’t run at press-release speed there.

Here’s the part I keep coming back to: the crowd didn’t wait for the lab. A community single-file build — MATLOWAI’s INT8 ConvRot bake — sits at 48,756 downloads, more than double the official preview. It folds the turbo and motion-smoothing LoRAs into one 21 GB ComfyUI file, and one partition serves both the first/last-frame and reference-conditioning paths — the two things FastVideo hasn’t distilled yet. And it ships measured timings on the builder’s hardware instead of marketing timings: a 10-second clip in 76 seconds at 4 steps on a 96 GB RTX PRO 6000; the same build takes 292 seconds at 25. Baking the LoRAs in saves about 21 GB of peak VRAM — 47.8 GB vs 68.9 GB on the same take — at the cost of never turning them down. So that’s the 4-step era on hardware a person can own: twelve times faster than fifty steps, and a workstation card still spends 76 seconds rendering ten seconds of video. Not instant.

What the four steps traded away is stated plainly on the card. Only the text-to-video-and-audio path is distilled — first/last-frame and reference-to-video are follow-up work — and the team admits preview quality sits below base H3 on hard motion, fine detail and some audio. Dense attention isn’t a drop-in either; the checkpoint is married to FastVideo’s sparse-attention kernels. That’s step distillation in general — train a student to match the teacher’s distribution in a handful of forwards instead of fifty — and it’s the same engine behind FastWan, LTX and the Qwen-Image Lightning LoRAs. It’s the cheapest speedup around because it multiplies across every generation you run, and the most fragile, because it’s tied to one checkpoint, one sampler, one kernel. One more asterisk: the whole thing inherits the MiniMax H3 Community License, not Apache-2.0 — read the territory and acceptable-use terms before pointing anything commercial at it.

Separate story, one paragraph: Google’s TimesFM-3, a 330M zero-shot forecaster that predicts whole multivariate horizons in a single forward pass and runs on a laptop or a CPU — and it moved 123,025 downloads last month, enormous for a forecasting model. Its top-of-table benchmarks are its own table. The real asterisk is the license: non-commercial. Your personal dashboards and research, yes; commercial self-hosting, no. If you want local forecasting, this is the one — read the license before you pipe it into anything you sell.

The most consequential release of the day isn’t a model. At IFA 2026, NVIDIA and Microsoft put one-click local model setup into the agent apps people actually use: Hermes Agent gets auto-detected GPUs and one-click local config through integrated llama.cpp on RTX and DGX systems; OpenClaw gets a Windows app that does the same for RTX cards with 24 GB or more; Perplexity’s Portable Computer comes to Linux first, Windows later. llama.cpp up to 1.9× throughput on a 5090, vLLM 1.2× on RTX PRO 6000 Blackwell and up to 1.4× across two DGX Sparks, flowing through Ollama and LM Studio too. And NVIDIA PAIR — free and open-source — routes inference jobs to whichever compatible machine on your network has room (Apple M4+, RTX 20-series+), so a subagent-heavy run spreads across your hardware instead of queueing on one GPU. The friction that kept local agents a weekend of configuration just dropped, in the harnesses people already run; the compact RTX Spark Windows boxes from Lenovo and Acer land in October.

The rest, plainly: the download charts are still another uncensored Qwen3.8-27B re-slice every day — the newest DavidAU build is at 174,405 downloads and nothing new in it. And Spark-X2.5-1.7B showed up as the pocket sibling of the 4B covered the day before; the 4B is still the one to pick.

Download and like counts verified via the HuggingFace API. FastH3 speed, quality and scope claims are FastVideo’s; the MATLOWAI timings are measured on the builder’s card; TimesFM-3 benchmarks are Google’s; the IFA items are from NVIDIA’s announcements.