Open Source AI Video Generators in 2026: What Actually Runs on Your Hardware

Introduction

"Open source video generator" is the most misleading phrase in AI video right now, and it costs people weekends.

Here is the pattern. You read that a new model is open source. You download something between 12 GB and 80 GB, install ComfyUI, and discover one of three things: the weights do not exist for the model you wanted, your GPU has enough memory for the model but not for the text encoder, or the license prohibits the commercial use you had planned.

I have hit all three. So this is the version I wish had been published: which video models genuinely publish downloadable weights in 2026, what each one needs on the hardware side, what the licenses actually permit, and — the part that decides the project — when running locally is the wrong choice entirely.

The short version up front: the newest Wan models are not open, the newest open Wan model is still Wan 2.2, and for most creators a hosted free tool beats a local install on cost until you are generating at real volume.

TL;DR

• Alibaba's open weights stop at Wan 2.2. Wan 2.5, 2.6, 2.7 and 3.0 are hosted endpoints on Alibaba Cloud Model Studio — no downloadable weights. Any page claiming otherwise is describing an release that did not happen.
• Wan 2.2 is still the safest default open model: Apache 2.0, no commercial restrictions, runs from 8 GB VRAM on the TI2V-5B GGUF variant with offloading.
• LTX-2.3 is the only open model with native synchronized audio in a single diffusion pass, and the only one reaching 4K at 50 fps locally.
• VRAM floor is not the whole story. The model may fit while the text encoder does not; budget 4–8 GB above the published minimum.
• License is the real gate, not hardware. Wan 2.2 and Mochi 1 are Apache 2.0; LTX-2.3 needs a commercial licence above $10M revenue; HunyuanVideo 1.5 caps at 100M monthly users.
• A 5-second 480p clip costs roughly $0.02–$0.03 on rented GPU time — cheaper than API video at volume, more expensive than a free hosted tier at hobby volume.

What "open source" means for a video model

Almost nothing in this space is open source in the software sense. Nearly all of it is open weight: the model file is downloadable, but training data and the full pipeline stay private.

Label	What you actually get	Commercial use
Truly open	Weights, training data, code, permissive licence	Usually yes
Open weight	Downloadable weights, inference code	Depends on licence
Hosted API	No download, pay per call	Usually yes, per terms
Demo only	Playground access, no export	No

For practical purposes, three questions decide everything: can I download it, can I run it on hardware I own or rent, and can I sell the output. A model that fails any of those is not open for your purposes, whatever the announcement said.

The open-weight video models that exist in 2026

These are the ones you can actually download and run today, with the hardware floors vendors and independent testing report:

Model	Smallest workable variant	Min VRAM	Max resolution	Native audio	Licence
Wan 2.2	TI2V-5B GGUF, offloaded	8 GB	720p	No	Apache 2.0
Wan 2.2 A14B MoE	GGUF	12 GB	720p	No	Apache 2.0
Wan 2.2 A14B MoE	BF16	40–48 GB	720p	No	Apache 2.0
LTX-2.3	Distilled NVFP4	16 GB	4K @ 50 fps	Yes	LTX-2 Community / Commercial
LTX-2.3 distilled FP8	FP8	32 GB	4K @ 50 fps	Yes	LTX-2 Community / Commercial
LTX-2.3 dev	BF16	80 GB	4K @ 50 fps	Yes	LTX-2 Community / Commercial
HunyuanVideo 1.5	FP8 + CPU text-encoder offload	14 GB	1080p	No	Tencent Community
HunyuanVideo 1.5	FP16	24 GB	1080p	No	Tencent Community
CogVideoX	2B FP16	16 GB	720p	No	Apache 2.0
CogVideoX	5B BF16	24 GB	720p	No	CogVideoX Licence
Mochi 1	FP8	20 GB	480p	No	Apache 2.0
Stable Video Diffusion	SVD-XT	8 GB	576×1024	No	Stability Community

Wan 2.2 — the default for a reason

Wan 2.2 shipped from Alibaba's Tongyi Lab in July 2025 under Apache 2.0, and it remains the newest Wan generation Alibaba has released weights for. Its Mixture-of-Experts design splits denoising between high-noise and low-noise experts, so detail improves without compute scaling linearly.

Two usable tiers. TI2V-5B handles text-to-video and image-to-video at 720p and, with a GGUF quantisation and memory offloading, starts around 8 GB of VRAM — consumer laptop territory. A14B MoE carries 27B parameters with 14B active per step and produces visibly better output; it wants 40–48 GB for the full-quality path, though quantised community builds run lower.

That combination — permissive licence, two hardware tiers, a large LoRA ecosystem — is why Wan 2.2 is still what most local video workflows are built on. It is also why open-weight Wan and hosted Wan have drifted apart as products.

If you want output today, start here: Launch Wan 2.2 Now →

LTX-2.3 — the one with real audio

LTX-2.3 generates synchronized audio and picture in a single diffusion pass, which no other open-weight video model in this comparison does. Everything else requires a separate audio step, and every separate audio step introduces drift you fix in an editor.

It also goes to 4K at 50 fps locally, which is a genuinely different output class. The costs are memory — 16 GB at the distilled NVFP4 build, 32 GB at FP8, 80 GB for dev BF16 — and licence: the LTX-2 Community licence is free for non-commercial use and for organisations under $10M revenue, with a commercial licence above that.

If your work needs sound-on video and you are set up for a 24 GB card plus offloading, this is the model to test first.

HunyuanVideo 1.5 — cinematic motion on one consumer card

Tencent's 1.5 lands around 14 GB with FP8 weights and CPU text-encoder offloading, and reviewers consistently put it at or near the top for photoreal faces and dynamic motion among open models. 1080p is reachable. The licence is the Tencent Community licence: free commercially up to 100M monthly active users.

The trade is ecosystem. Fewer community LoRAs, fewer battle-tested ComfyUI graphs than Wan 2.2.

CogVideoX — best prompt adherence

CogVideoX-5B holds multi-clause, structured prompts together better than the alternatives, which matters when your prompt is a shot list rather than a vibe. The 2B variant runs on 16 GB under Apache 2.0; the 5B needs 24 GB and its own licence. Fine-tuning is supported through the CogKit framework.

Mochi 1 — the fine-tuning base

Mochi 1's official LoRA trainer runs on a single GPU, which makes it the most accessible base in this list for training your own style or character. It tops out at 480p, so it is a training and prototyping tool rather than a delivery model.

Stable Video Diffusion — the 8 GB fallback

SVD and SVD-XT still run on 8 GB and slot into existing Stable Diffusion and AnimateDiff workflows. Quality is behind the 2026 field; convenience is not.

The VRAM trap nobody mentions

Published VRAM floors assume the model alone. In a real ComfyUI or Diffusers pipeline you also load a text encoder, a VAE, and usually a second pass.

What you think needs 8 GB	What actually happens
Model weights	Fit
Text encoder	Pushes you into offloading, 2–4× slower
VAE decode at final resolution	Spikes above steady-state memory
OS + display	1–2 GB gone before ComfyUI starts

Ready to try it yourself? Try Wan 2.2 Free →

Rule of thumb from running these: add 4–8 GB to the floor, or accept offloading and the slowdown it brings. The 8 GB Wan 2.2 claim is true — with GGUF quantisation and sequential offloading — and the difference between it and a 24 GB card is the difference between four minutes and forty seconds per clip.

Licences decide more than hardware

Licence	Commercial use	Models
Apache 2.0	Unrestricted	Wan 2.2, Mochi 1, CogVideoX-2B
LTX-2 Community / Commercial	Free below $10M revenue, licence above	LTX-2.3
Tencent Community	Free below 100M MAU	HunyuanVideo 1.5
CogVideoX Licence	Permitted with restrictions on prohibited uses	CogVideoX-5B
Stability Community	Permitted with conditions	SVD

Two practical notes. Apache 2.0 is the only licence here that needs no reading — if commercial certainty matters more than marginal quality, Wan 2.2 and Mochi 1 are the safe picks. And licence terms get revised; the version you agreed to when you downloaded is not necessarily the version shipping today, so check before a paid deployment.

What local generation actually costs

A 5-second 480p clip on an RTX 4090 takes roughly 4–6 minutes with Wan 2.2's 5B variant. On a rented A6000 48 GB at around $0.35/hour, that works out near $0.02–$0.03 per clip — an order of magnitude cheaper than API video at volume.

The arithmetic flips at low volume and at iteration-heavy work:

Your volume	Local	Hosted
1–10 clips to test an idea	Bad fit: hours of setup for minutes of output	Free tier, minutes
A few dozen clips a month	Marginal: beats subscriptions only if you already own the GPU	Comparable
Hundreds of clips, stable prompts	Clearly cheaper, once the pipeline exists	Per-call costs dominate
Anything needing audio	LTX-2.3 only, and 32 GB of VRAM	Hosted models with native audio
Work you cannot upload	Required regardless of cost	Unusable

The hidden line item is iteration time. Local generation is not just compute — it is node-graph debugging, dependency pinning, and re-downloading a 30 GB checkpoint because a custom node broke. That is fine if you enjoy it, expensive if you bill hourly. For a closer look at how it stacks up against other models, see Wan 2.5 vs Kling 3.

A workflow that survives contact with reality

Step 1 — Decide where the constraints are. Commercial use, raw footage you cannot upload, and volume above a few hundred clips are the three conditions that justify local. Otherwise start hosted.

Step 2 — Prototype on a free hosted tool. Get the prompt and the motion direction right before you spend an evening on installation. A composition that fails at 480p on a hosted model fails locally too.

Step 3 — Pick the model by licence, then by VRAM. Apache 2.0 and 8 GB points at Wan 2.2 TI2V-5B. Native audio points at LTX-2.3 and 16 GB minimum. Multi-clause prompts point at CogVideoX.

Step 4 — Prove one clip end to end before batch work. Quantisation, offloading and resolution all interact. Confirm a single 5-second output, with audio if required, before writing a batch loop.

Want to see the difference on your own footage? Start creating with Wan 2.2 →

Step 5 — Keep a hosted fallback for the shots that fail. Local models have specific failure modes — hands, fast lateral motion, long durations. Re-rolling the same failing shot locally forever is a worse use of time than paying for one hosted render.

Steps 2 and 5 are why I keep the free Wan tools open in a tab even when a local ComfyUI graph is running. They are a cheap way to answer "is this prompt even viable" before touching a checkpoint.

When to skip local entirely

Open weights solve a licensing and privacy problem. They do not solve a quality problem, and they are not the fastest route to a finished clip.

Situation	Better choice
Testing whether an idea works at all	Hosted free tier — text to video
Animating a product photo for an ad	Hosted image to video, no install
Client work needing the newest model	Hosted Wan 2.7 — no weights exist locally
30-second narrative shots	Hosted; open models cap shorter
Data you legally cannot upload	Local, non-negotiable
Fine-tuning a style or character	Local, on Wan 2.2 or Mochi 1

Most creators end up in the top four rows, which is the honest argument for hosted tools: the open-weight route is a licensing decision, not the default path to better video.

Try open-weight Wan without downloading 30 GB

If the reason you looked up "open source AI video generator" was to get away from paywalls and watermarks rather than to own a checkpoint, you can skip the installation and use the same model family in a browser.

• Free Wan video generator — runs the Wan family with no GPU, no ComfyUI, no checkpoint downloads
• Image to video — animate a single still without building a node graph
• Text to video — the fastest way to test whether a prompt holds
• No watermark on paid output, no queue behind a 40 GB download
• When you need the newest model, the hosted route covers Wan 2.7 and Wan 3.0, which have no local weights at all

If you do outgrow it, the local path is still there — Wan 2.2 under Apache 2.0, unchanged. The point is to reach that decision with a working clip in hand rather than a half-finished install.

FAQ

Is Wan 2.7 open source?

No. Alibaba's open-weight releases stop at Wan 2.2 (T2V-14B, I2V-14B, TI2V-5B) on Hugging Face. Wan 2.5, 2.6, 2.7 and 3.0 are served as hosted endpoints through Alibaba Cloud Model Studio and resellers. Pages describing Wan 2.7 weights under Apache 2.0 are describing a release that did not happen.

What is the best open source AI video generator in 2026?

Wan 2.2 for most people: Apache 2.0, an 8 GB entry point on the 5B GGUF, and the largest LoRA ecosystem. LTX-2.3 if you need native synchronized audio or local 4K. HunyuanVideo 1.5 for cinematic motion on a single consumer card.

How much VRAM do I need?

8 GB for Wan 2.2 TI2V-5B with GGUF and offloading, 12 GB for the A14B MoE quantised, 14 GB for HunyuanVideo 1.5 with FP8 and CPU offload, 16 GB for LTX-2.3 distilled, 20 GB for Mochi 1 at FP8, 24 GB for CogVideoX-5B. Add 4–8 GB for the text encoder, VAE and ComfyUI overhead in practice.

Which open model generates audio?

LTX-2.3 is the only one in this comparison that produces synchronized audio and video in a single pass. Wan 2.2, HunyuanVideo 1.5, CogVideoX and Mochi 1 are silent and need a separate audio step.

Can I sell videos generated with these models?

Wan 2.2, Mochi 1 and CogVideoX-2B are Apache 2.0 with no commercial restrictions. LTX-2.3 requires a commercial licence above $10M annual revenue. HunyuanVideo 1.5 permits commercial use below 100M monthly active users. Verify the current licence text before deploying commercially.

Is local generation cheaper than hosted?

Only above a certain volume or when privacy forces it. A 5-second 480p clip costs roughly $0.02–$0.03 on rented GPU time, which beats API pricing at hundreds of clips a month — but setup time, debugging and re-downloads dominate below that.

Can I run these on a laptop?

Wan 2.2's 5B GGUF at 8 GB and SVD at 8 GB are the two realistic laptop options, both with offloading and both slow. Expect minutes per short clip, not seconds.

Related guides

• Wan video models compared — every Wan generation side by side
• Wan 3.0 complete guide — what the hosted flagship does that open weights do not
• Z-Image Turbo FP8 local guide — the local image counterpart if you are building an offline pipeline
• Free Wan 2.2 unlimited guide — how the free hosted tier compares to running 2.2 yourself

References

• Best open-source AI video generation models (2026) — VRAM floors, per-variant hardware requirements, licence comparison and the ~$0.35/hour A6000 cost basis used above
• Wan-AI on Hugging Face — the official weight repository, which stops at Wan 2.2
• Wan2.2 on GitHub — Apache 2.0 release, TI2V-5B and A14B MoE variants, model download commands
• LTX-2.3 model card — single-pass synchronized audio, 4K at 50 fps, and the distilled/FP8/dev memory tiers
• LTX-2 licence — the $10M revenue threshold for commercial use
• HunyuanVideo 1.5 — FP8 and CPU-offload configuration, and the Tencent Community licence terms
• CogVideo — CogKit fine-tuning support and the 2B/5B licence split
• Mochi 1 fine-tuning README — single-GPU LoRA training
• Wan 2.7 and 3.0 model listings — the hosted Wan endpoints, showing the API-only distribution of post-2.2 generations
• Wan 2.7 AI Video Generator on Pollo.ai — subscription access to the hosted model, with fewer restrictions than a free tier