← Back to blog
AiAbout 6 min read

Two Open Video Models Lost Weight This Week. Neither Published the Column That Matters.

Published Nov 8, 2026
Two Open Video Models Lost Weight This Week. Neither Published the Column That Matters.

On November 7, Video Rebirth published HyperFlow H3 Flash and put it on Hacker News. On roughly the same day, Vimoo-AI shipped ComfyUI weights for Vimoo 1.0 Lite. Both are open video models. Both are dramatically smaller than the thing they came from. Both arrive with speed tables.

Both are missing the same column.

What Video Rebirth cut

HyperFlow H3 Flash is a compressed, few-step version of MiniMax-H3, the open model that generates video and synchronised stereo audio together. The transformer drops from roughly 31 billion parameters to 13.7 billion. Sampling runs in four, six or eight steps.

An engineer on the project, posting as jiant, listed three changes. They pruned three of the model's fifty blocks. They factored 102 linear layers into low-rank form. They applied 8-bit integer quantisation with a Hadamard rotation to 180 linear layers.

The fourth change is the interesting one, because it is not compression at all. They added a two-value conditioning scheme so each sampling step knows both the current noise level and where it should land. That is what lets the model run in four to eight steps without classifier-free guidance, which is normally the thing that doubles your inference cost.

The numbers, as reported by the authors: a 5.2-second clip at 1344×768 in six steps takes 8.7 seconds on four B200 GPUs, 36 seconds on a single H200, and 92 seconds on a single RTX 5090. NVIDIA's SOL-H3 inference engine, built by a partner, reportedly runs it in real time on four Blackwell-class GPUs. Output is 5.2 to 15 seconds at 24 frames per second.

For workstations there are GGUF builds from 7.2GB at Q3_K_M to 15.2GB. The README recommends the 9.1GB Q4_K_M build for 24GB cards, and says it runs on 24GB and 32GB GPUs in machines with 16GB of system RAM. Sixteen-gigabyte GPUs are listed as not yet measured, which is an unusually honest thing to leave in a README.

A tall stack of blank brushed aluminium canisters on a concrete shelf, three removed.

Two practical details the speed table hides. The first run of the Python package downloads about 90GB, or about 68GB with an FP8 text encoder. Kernels compile for roughly a minute. Neither is a dealbreaker. Both mean the model does not fit in the "just try it" category of software.

What Vimoo cut

Vimoo 1.0 Lite is about 9.5 billion parameters in a single model that handles both text-to-video and image-to-video, plus a distilled LoRA adapter for the same model. It ships with text encoders, a vision encoder, a VAE, ComfyUI nodes and workflow templates.

The licence is Apache 2.0, which is the part most readers will care about more than the benchmark. Apache 2.0 means commercial use without a territory clause, without a research-only rider, and without the email-the-authors step that a lot of open video releases bury in a model card.

The headline number: at 1280×720 and 121 frames, the distilled model generates video in 9.7 seconds. The comparison offered is Wan 2.2 at 490 seconds with CFG enabled.

That comparison is doing a lot of work and deserves to be handled carefully. A distilled model and a CFG-enabled base model are not the same measurement. One is a student that has been trained to skip the guidance step. The other is a teacher running the full procedure. Comparing them tells you that the distilled path is faster, which everybody already knew. It does not tell you how much quality the student gave up.

The full Vimoo 1.0 is a hybrid design on a Dual-Stream MMDiT with 11.7 billion parameters, a video VAE at 4×16×16 compression with 48 latent channels, Qwen2.5-VL-7B for prompt understanding, and Glyph-ByT5-v2 for typography. Vimoo Enhancer handles upscaling to 1080p. A five-second 720p clip at 24fps takes about five seconds, which is a different claim than the Lite distilled number and is worth keeping separate in your head.

The missing column

Here is the shape of the week. Three open video models shipped compressed: HyperFlow H3 Flash, Vimoo 1.0 Lite, and Pika H3 Flex before them. Each one tells you parameters removed, steps removed, seconds measured, gigabytes required. None of them tells you what happened to the output.

The Hacker News thread on HyperFlow made the objection directly. One commenter compared it to announcing a JPEG compressed to half its size without checking what it looks like. Another asked for a head-to-head against the original model. Someone pointed at the project's sample gallery, which is the standard answer and the standard problem: curated clips demonstrate that the model can output good frames, not how often it does.

A third commenter asked how it compares with FastVideo's FastH3 variants, which they run on a laptop-class card. No answer from the authors appeared in the part of the thread that was public.

This is the gap that matters for anyone deciding whether to adopt one of these. The compression techniques are all reasonable. Structured pruning, low-rank factorisation, INT8 with rotation, distillation, step-aware conditioning: each has a known cost profile in image models, where quality can be measured per-image with FID or GenEval or a human side-by-side. Video is harder. Temporal consistency, motion coherence and audio sync do not have a single number that survives compression well.

So the field defaults to the numbers that are easy to publish. Parameters are easy. Seconds are easy. Quality is a reviewer, a blind test, and a weekend.

How to read a release like this

If you are evaluating one of these for actual production work, the speed table is a ceiling, not a plan. Run ten prompts you already know through your current service and through the 9.1GB build, then have someone score them blind. That takes an afternoon and produces the column the release left out.

Check both licences before you ship anything commercial. HyperFlow is a derivative of MiniMax-H3, so it inherits constraints from the base model's licence, and the base licence is not Apache 2.0. Vimoo Lite is the simpler case, because Apache 2.0 is Apache 2.0.

And keep the disk space in your budget. A 90GB first-run download is trivial on a server and annoying on a laptop, and it is the kind of detail that turns a five-minute evaluation into a two-hour one.

The compression race is real and it is useful. A video model that runs on a single 24GB card is a different product from one that needs a node. But the race is being scored on the parts that are cheap to measure, and the part that decides whether you can use it is the part nobody put in the table.

Update your own numbers. That is the only table that ends up being true for your prompts.

Related articles