Best Local Video Generation Models in 2026: Open Weights by VRAM, Mac and Licence

For most teams the best local video generation model in November 2026 is still Wan 2.2: it is Apache 2.0, ComfyUI says its 5B model should fit an 8 GB card with offloading, and its 14B experts are 14.29 GB each in FP8. For sound generated with the picture, LTX-2.5 is free below $10 million in annual revenue and MiniMax H3 below $20 million, though H3's licence excludes the US, EU, UK and South Korea. Ollama does not generate video.

Local video generation changed shape in 2026. Several open models released this year generate audio alongside the frames, accept reference images and clips, and arrive with quantised files for consumer cards, while Alibaba’s newest Wan versions stayed behind APIs. This guide covers text-to-video, image-to-video and video-to-video models with downloadable weights, checked against model cards, licence files, file listings and runtime documentation on 4 November 2026. Sizes are published file sizes in gigabytes, and memory figures that are our own estimates say so.

Still pictures need different models, covered in our guide to open image generators by graphics memory.

Quick picks by VRAM and platform

The open-weight video models that matter in November 2026

Output differs as much as size. Wan 2.2’s 14B models generate 480p and 720p, and the 5B model 1280 × 704 at 24 fps. LTX-2.5’s default two-stage workflow upscales a 768 × 512 base to 1536 × 1024, reaches 4K with its upscalers and, according to ComfyUI, runs at up to 50 fps. MiniMax H3 makes 4 to 15 seconds at a 768-pixel short edge, HunyuanVideo 1.5 renders 480p or 720p with an optional pass to 1080p, LongCat-Video targets 720p at 30 fps, and MAGI-2-preview makes only 10-second clips, refined to 1080p.

Each model also needs a text encoder, and some encoders rival the video model in size: Qwen3-VL-32B for H3 (15.69 GB in NVFP4), Gemma 4 12B for LTX-2.5 (15.37 GB in INT8), Qwen2.5-VL 7B for HunyuanVideo 1.5 (9.38 GB in FP8) and umT5-XXL for Wan 2.2 (6.74 GB in FP8). ComfyUI runs the encoder once per prompt and can move it to system memory before sampling, so the video model sets the VRAM floor while the encoder mostly costs RAM and disk.

Two more 2026 releases are worth knowing. Wan-Animate-2, released on 7 August 2026 under Apache 2.0, animates a character image with the motion of a driving video. LingBot-Video from Robbyant arrived on 9 July 2026 under Apache 2.0, as a 30B-A3B mixture-of-experts model and a 1.3B dense model aimed at embodied-intelligence research. Older releases such as Genmo’s Mochi 1 (October 2024) and CogVideoX (August 2024) are still on Hugging Face.

Some well-known video models cannot be downloaded at all. Kling, Seedance, Veo and Runway are hosted services, and so are Alibaba’s Wan releases after 2.2: on 4 November 2026 neither the Wan-AI organisation on Hugging Face nor Wan-Video on GitHub published weights for Wan 2.5 or later, and ComfyUI offers Wan 3.0 only as an API node.

Licences: which local video models you can use commercially

• Apache 2.0 or MIT, commercial use allowed: the Wan 2.2 family including Wan-Animate-2, Kandinsky 6.0, LongCat-Video, LingBot-Video and MAGI-2-preview. The Wan 2.2 README adds that its authors claim no rights over generated content.
• Free below a revenue line: LTX-2.5 uses the LTX-2.x Community License, dated 11 August 2026 and amended on 29 October 2026, and LTX-2.3 the earlier LTX-2 Community License. Both set a $10 million annual-revenue line above which an entity needs a paid Commercial Use Agreement. Under the 2.x text, testing, evaluation and non-commercial research stay free, while training or distilling a model for commercial use counts as commercial. The 2.x licence also obliges users to keep watermarking, provenance and disclosure features intact and to follow AI rules such as the EU AI Act.
• Limited by territory: MiniMax H3’s licence excludes the United States, the European Union, the United Kingdom and South Korea, and requires written authorisation above $20 million of yearly revenue. HunyuanVideo 1.5’s Tencent Hunyuan Community License does not apply in the EU, the UK or South Korea, and licensees with more than 100 million monthly active users need a separate licence from Tencent.

Read the licence text itself rather than a model-card summary, and treat this section as general information rather than legal advice. Our open-weight licence checklist for legal teams covers the review, and for clips published in the EU our explainer on the AI Act’s transparency duties covers how synthetic video must be marked and labelled.

Best local video generation model for 8 and 12 GB GPUs

Small cards depend on offloading, which keeps most weights in system RAM and streams them to the GPU as needed, so plan RAM as carefully as VRAM.

• Wan 2.2 TI2V-5B is the safest pick. ComfyUI’s guide says it should fit on 8 GB with native offloading; the FP16 model file is 10.00 GB and the FP8 encoder 6.74 GB. One model handles text-to-video and image-to-video under Apache 2.0.
• Kandinsky 6.0 Lite adds synchronised audio under MIT. Its transformer is 7.48 GB and its Diffusers examples enable CPU offloading, so a 12 GB card is a reasonable target by our estimate; the model card gives no figure.
• MiniMax H3 reaches these cards through WanGP, whose author reports 15 seconds of 480p video in 5 to 6 GB, or through GGUF builds from 6.72 GB. Our step-by-step H3 guide for ComfyUI covers both routes and the licence’s territory limits.

Tencent’s README also lists WanGP as running HunyuanVideo 1.5 in as little as 6 GB. Prices and memory for each card class are in our guide to GPUs for local AI.

Best local video generation model for 16 and 24 GB GPUs

At 16 GB, HunyuanVideo 1.5 fits the class it was built for: Tencent lists 14 GB as the minimum with offloading, and ComfyUI’s FP8 builds are 8.33 GB per model. It renders 480p or 720p and adds a distilled super-resolution model for 1080p, but its licence does not cover the EU, UK or South Korea. Wan 2.2’s A14B models run as two experts of 14.29 GB each in FP8, a high-noise one for early steps and a low-noise one for later steps, and ComfyUI’s guide pairs them with LightX2V’s 4-step LoRAs to cut sampling time.

At 24 GB, Wan 2.2 TI2V-5B runs at full precision; Wan’s README gives an RTX 4090 as its example card, with offload flags. ComfyUI’s HunyuanVideo 1.5 guide targets this class for 720p. MiniMax H3’s default ComfyUI files suit it too: a 20.97 GB INT8 model, and an SGLang run on one RTX 4090 that peaked near 18 GB. Lightricks publishes INT8 “lower-VRAM” files for LTX-2.5, a 21.50 GB transformer and a 15.37 GB encoder, yet still lists 32 GB as the minimum, so a 24 GB card is below its specification and depends on offloading.

32 GB, 48 GB and server-class GPUs

32 GB is the memory of both the RTX 5090 and AMD’s Radeon AI PRO R9700, and it is LTX-2.5’s stated minimum. It also holds one of Wan 2.2’s 16-bit experts at 28.58 GB. On AMD hardware, WanGP supports RDNA 2 to 4 GPUs and MiniMax H3’s vLLM-Omni recipe covers ROCm; our guide to the Radeon AI PRO R9700 covers that card for local AI.

48 to 96 GB workstation cards take the larger checkpoints at 16-bit: MiniMax H3’s pruned model (40.23 GB), LTX-2.5’s transformer (42.02 GB) and, on a 96 GB card, Kandinsky 6.0 Pro’s 69.73 GB transformer, each with its encoder swapped out by our estimate. Lightricks recommends an 80 GB A100 or an H100 for LTX-2.5.

Servers. Wan 2.2’s single-GPU A14B script needs an 80 GB GPU unless you add its offload flags, MiniMax H3’s SGLang reference runs on four to eight data-centre GPUs, and MAGI-2-preview requires eight Hopper GPUs. Wan and MiniMax both document multi-GPU sequence parallelism for shared generation servers.

On unified-memory PCs such as NVIDIA’s RTX Spark laptops, what matters is how much of the shared memory Windows assigns to the GPU, which our RTX Spark guide works through for each configuration; video models draw on the same budget.

LTX-2.5 vs MiniMax H3

Choose LTX-2.5 for higher resolution and frame rates, a worldwide licence and its IC-LoRA tools, provided your organisation stays under the revenue line or buys a licence. Choose MiniMax H3 for reference-driven work that mixes images, clips and voices, if you operate outside its excluded territories. In ComfyUI, LTX-2.5 has native text-to-video, image-to-video, first-last-frame and DFR templates. Neither vendor publishes ComfyUI timings, and this guide does not rank their visual quality.

Best local video generation model for Mac

Macs pool CPU and GPU memory, and our Mac model guide treats about 80% of it as usable. Video adds a second constraint: Lightricks requires an NVIDIA GPU with CUDA for its own LTX-2.5 pipelines, and Comfy-Org’s INT8 ConvRot files are meant for CUDA builds of PyTorch. Two Apple-silicon routes work:

• Draw Things. Its open-source engine includes implementations of Wan, LTX-2, HunyuanVideo, Kandinsky and MiniMax H3, and its October 2026 benchmark generated H3 video at 1024 × 512 on Macs from a 24 GB M4 Pro to a 256 GB M5 Ultra, using quantised builds.
• mlx-video. This MIT-licensed project runs LTX-2 and LTX-2.3 with audio, and Wan 2.1 and 2.2 (T2V-14B, TI2V-5B and I2V-14B), on Apple’s MLX framework. Its last commit was on 13 May 2026, before LTX-2.5 and H3 appeared.

The 14B and 22B models benefit most from memory, and our Mac Studio M5 Ultra guide covers Apple’s largest desktop configurations.

Runtimes: ComfyUI, Diffusers, WanGP and Ollama

• ComfyUI (version 0.44.0 on 4 November 2026) has official templates for Wan 2.2, Wan Animate 2, LTX-2.5, MiniMax H3 and HunyuanVideo 1.5 in its Template Library.
• Diffusers 0.40.0, released on 20 August 2026, added LTX-2.5, MiniMax H3 and Wan-Animate-2; Wan 2.2 has had Diffusers checkpoints since July 2025, and Kandinsky 6.0 is published in Diffusers format.
• WanGP is a web app for low-memory GPUs that runs Wan, MiniMax H3, LTX-2 to 2.5, HunyuanVideo 1 and 1.5, LongCat and Kandinsky. Its README says select models run in as little as 4 GB of VRAM, and it supports AMD RDNA 2 to 4.
• stable-diffusion.cpp runs GGUF builds of Wan 2.1 and 2.2, MiniMax H3, LTX-2.3 and 2.5, HunyuanVideo 1.5 and LingBot-Video from C++.
• Ollama is not an option for video. Its release notes through version 0.40.14, published on 3 November 2026, mention no video generation, and the experimental image feature it added earlier in 2026 was withdrawn in 0.32.6.

A practical local video generation setup:

1. Pick the model by licence first, then by memory, using the tables above.
2. Install and update ComfyUI, or WanGP on a small card.
3. Open Template Library, choose Video and load the model’s template; for LTX-2.5, request access on Hugging Face first.
4. Budget disk from the published files: about 37 GB for Wan 2.2 A14B in FP8 with its encoder and VAE, about 40 GB for LTX-2.5’s lower-VRAM set, about 42 GB for H3’s template and about 21 GB for HunyuanVideo 1.5 in FP8.
5. Start at the template’s default size and length, then raise both once a run completes, because activation memory grows with each. Weight arithmetic is parameters × bits ÷ 8, the same method our VRAM calculator applies to language models.

Published generation times are sparse and use different settings. Wan says TI2V-5B makes a 5-second 720p video in under 9 minutes on one consumer GPU. Tencent says its step-distilled 480p image-to-video model finishes within 75 seconds on one RTX 4090. MiniMax’s guide cites 559.67 seconds for a 5-second 1344 × 768 H3 clip at 50 steps on two RTX 5090s. The Lightricks model card and documentation pages we checked give no timings.

How VDF AI fits

VDF AI does not generate video, and none of the models in this guide run inside it. It is a private AI platform for language-model work: governed agents, VDF AI Networks for multi-agent orchestration, private RAG over company documents, role-based access control on every plan and audit logs.

Where a team already runs open models on its own servers, VDF AI Router can register those language-model endpoints, for example a vLLM or Ollama server, next to approved cloud models, and air-gap mode keeps every request on the local ones.