Why we sometimes pick older models

The setup

Every AI model is a trade-off. Faster but worse at faces. Better at motion but heavier on memory. Free locally but caps at 5 seconds. Cloud-only but the cleanest face lock anywhere.

Most AI tools hide that trade-off. One Generate button, no idea which model is firing or why. The tool picks for you, and you find out it was the wrong pick when the output looks weird. Hybrig flips that — every model has a public card on /models with what it’s good at, what it fails at, and when to swap to a different one. Including older ones.

Worked example: Wan 2.1 vs Wan 2.2 vs Wan 2.2 FLF2V

Three local video models, all in the “Wan” family from Alibaba. All run on your own GPU. All free. Three different jobs. Here’s how to pick.

Job 1: Standard image-to-video (one starting frame, just add motion)

Winner on a 4090: Wan 2.1. The full Wan 2.2 model is too big to fit on most consumer GPUs in its uncompressed form. The shrunk-down version (sometimes called “Turbo,” sometimes called a quantized build) runs fine on a 4090 but scores worse than Wan 2.1 in our tests at the same number of refinement passes. Newer model, same hardware, worse output.

That’s the lesson hiding under “newer is not always better.” The model improved on paper. Your hardware is the same. The compressed version that fits your hardware doesn’t carry the on-paper improvement. So 2.1 wins.

Job 2: First/last frame interpolation (morph between two stills)

Winner: Wan 2.2 FLF2V. Plain English: feed it two images — a starting frame and an ending frame — and it generates the video that morphs between them. Perfect for weathering time-lapses (fresh shingle year 0 → aged shingle year 25), before/after restoration shots, age progression on a locked portrait. Anywhere you want both ends of the clip locked to specific images.

Wan 2.1 simply doesn’t have this capability. It can’t lock an ending frame — it just guesses where the motion goes. So for this job, you must use Wan 2.2 FLF2V. Hunyuan and LTX don’t have it either. The newer model wins here because the capability itself is new.

The honest summary

Two adjacent jobs, two different right answers. If you reflexively picked “the latest Wan” for Job 1, you’d ship worse output than going back to 2.1. If you reflexively picked the most-recent default for Job 2, you’d burn cloud credits on Kling because you didn’t know your local stack had the capability all along. Picking by job, not by version number, is the whole skill.

Plain-English glossary

Every model card on Hybrig avoids jargon when it can, and translates it when it can’t. Quick reference:

• Steps (sometimes “sampler steps”) — how many times the model refines the image before delivering. More steps = more polish, more time. 30 is a common middle.
• VRAM — the dedicated memory on your graphics card, separate from regular RAM. Big models need a lot of it. A 4090 has 24 GB; a 3060 has 12 GB.
• Quantization (e.g. Q5_K_M, Q4_K_M, fp8) — a compressed version of the model that runs on smaller GPUs at slightly lower quality. The number tells you how aggressive the compression is — lower number, smaller file, more quality lost.
• I2V — image-to-video. One starting image plus a motion prompt becomes a short clip.
• FLF2V — first/last frame to video. Two images (start + end) plus an optional prompt becomes a clip that morphs between them. Only Wan 2.2 has this locally.
• LoRA — a small extra training file that teaches a base model one specific thing (your face, a brand, a style). Pairs with a base model; doesn’t replace it.

How to read a model card on /models

Every card on /models has three sections that matter:

• Best for — the jobs this model wins on. If your job matches, you picked right.
• Fails at — the jobs this model loses on, AND when to swap to a different (sometimes older) model. This is the section most tools hide.
• Pairings — which other models stack well with this one. Hybrig is a workflow product; rarely does one model do everything alone.

When a model is added, removed, or swapped, those three sections update at the same time as the Studio palette node descriptions and the worker code. One source of truth, no drift between what the marketing page promises and what the Generate button does.

The bigger point

Hybrig is built around a workflow doctrine: the models are commodities, the chain is the product. That means picking the right tool for each link of the chain matters more than picking “the newest one” everywhere. Sometimes the right tool is two years old. Sometimes the right tool was released last week. The job tells you which.

We tell you which, in writing, in public, on every model card — and we update it when the call changes. That’s the editorial authority other tools can’t fake.