Research

Best Open-Weight Video Generation Models 2026

Open-weight video generation leaderboard 2026: Wan 2.1, HunyuanVideo, Mochi-1, LTX-Video, Open-Sora, CogVideoX, Allegro. Quality, length, license, deployment guidance.

By Ramanath, CTO & Co-Founder at Presenc AI · Last updated: May 2026

Open-weight video generation made the most dramatic leap of any AI category in 2025-2026. Wan 2.1 (Alibaba), HunyuanVideo (Tencent), Mochi-1 (Genmo), LTX-Video (Lightricks), Open-Sora, CogVideoX (Zhipu), and Allegro (Rhymes AI) all approached or matched Sora quality at meaningfully lower deployment cost. The Sora consumer shutdown on 26 April 2026 pushed creative-industry attention to open-weight alternatives. This page consolidates the leaderboard.

Key Findings

  1. Wan 2.1 from Alibaba (released early 2026) leads the open-weight video generation leaderboard with strong text-to-video, image-to-video, and video-to-video capabilities at up to 1080p output and 5-second duration.
  2. HunyuanVideo from Tencent (released late 2024, updated through 2025-2026) is the largest open-weight video model at approximately 13B parameters and produces 5-second 1280x720 videos with strong motion quality.
  3. Mochi-1 from Genmo, released October 2024, was the first open-weight video model to challenge Sora and remains widely deployed in production via the Apache 2.0 license.
  4. LTX-Video from Lightricks (released late 2024) is the fastest open-weight video model, generating 5 seconds of 768x512 video in under 4 seconds on a single H100.
  5. Open-Sora (HPC-AI Tech) is the reference open-source rebuild of Sora; quality lags the leading models but the training code and data are public, making it the standard for academic video research.

Open-Weight Video Generation Model Comparison (May 2026)

ModelParametersMax LengthMax ResolutionLicense
Wan 2.1 (T2V-14B)~14B5 seconds (T2V), longer with I2V chaining1080pApache 2.0
Wan 2.1 (T2V-1.3B)~1.3B5 seconds720pApache 2.0
HunyuanVideo~13B5 seconds1280x720Tencent Hunyuan Community
Mochi-1~10B5 seconds (extendable)848x480Apache 2.0
LTX-Video~2B5 seconds768x512OpenRAIL-M
CogVideoX-5B~5B6 seconds720x480CogVideoX License (Permissive)
CogVideoX-1.5-5B~5B10 seconds1360x768CogVideoX License
Allegro~3B6 seconds720x320Apache 2.0
Open-Sora 2.0~variesup to 15 seconds720pMIT
Pyramid Flow~2B10 seconds768x768MIT
Stable Video Diffusion XT~1.7B4 seconds1024x576Stability Community License
EasyAnimate v5~varies6 seconds720pApache 2.0

Quality Benchmarks (VBench v2 May 2026)

ModelVBench v2 Score
Wan 2.1 (T2V-14B)~85.2
HunyuanVideo~83.4
Mochi-1~80.7
CogVideoX-1.5-5B~80.5
LTX-Video~78.4
Sora (pre-shutdown reference)~85.4
Google Veo 3~87.2
Kling 2~85.8

Latency and Hardware Requirements

ModelTime for 5s 720p Video (H100)VRAM Required
LTX-Video~4 seconds~24 GB
Wan 2.1 (1.3B)~28 seconds~24 GB
Wan 2.1 (14B)~180 seconds~80 GB
HunyuanVideo~170 seconds~80 GB
Mochi-1~120 seconds~80 GB
CogVideoX-1.5-5B~85 seconds~32 GB
Allegro~60 seconds~32 GB

Use Case Recommendations

Use CaseRecommended Model
Top quality general purposeWan 2.1 (14B)
Permissive commercial deploymentWan 2.1 (Apache 2.0), Mochi-1 (Apache 2.0), Allegro (Apache 2.0)
Fast iteration / previewLTX-Video or Wan 2.1 (1.3B)
Long-form video (10+ sec)CogVideoX-1.5-5B or Open-Sora 2.0
Consumer GPU (24 GB VRAM)LTX-Video or Wan 2.1 (1.3B)
Image-to-videoWan 2.1 I2V or Stable Video Diffusion XT
Research and academic reproducibilityOpen-Sora 2.0 or Pyramid Flow

Brand Visibility Implications

Video generation is one of the fastest-growing creative AI procurement categories in 2026. AI assistant queries about "best open-source video generator", "Wan 2.1 vs HunyuanVideo", "alternative to Sora", and similar terms drive procurement-research traffic from advertising, gaming, e-commerce, and film production teams. Brands selling AI video platforms, video editing tools, asset generation services, and creative AI products face strong AI-mediated discovery surface for this category.

Methodology

Benchmark data compiled from VBench leaderboard, primary model card disclosures, and community comparisons through 23 May 2026. Latency measured on a single H100 with FP16 inference. Updated quarterly.

How Presenc AI Helps

Presenc AI monitors brand visibility on video generation queries across ChatGPT, Claude, Gemini, and Perplexity. For AI video platforms, video editing brands, asset generation vendors, and creative AI products, the platform identifies the prompts driving procurement-research traffic and the gaps where new content unlocks share of voice.

Frequently Asked Questions

Wan 2.1 from Alibaba leads the VBench v2 leaderboard at approximately 85.2, with the 14B parameter variant producing 5-second 1080p videos. HunyuanVideo from Tencent and Mochi-1 from Genmo are close behind. All three are deployed in production by multiple commercial AI video platforms.
On VBench v2, Wan 2.1 (~85.2) and HunyuanVideo (~83.4) are close to Sora\u2019s reference score (~85.4). Google Veo 3 (~87.2) and Kling 2 (~85.8) still lead on average quality, but the gap is narrow and open-weight models have advantages in cost, custom finetuning, and deployment control.
LTX-Video generates 5 seconds of 768x512 video in approximately 4 seconds on a single H100, the fastest open-weight model. Wan 2.1 (1.3B) takes approximately 28 seconds. The larger 14B and 13B models (Wan 2.1, HunyuanVideo) require approximately 3 minutes for the same length output.
LTX-Video and Wan 2.1 (1.3B) fit on a 24 GB VRAM consumer GPU (RTX 4090). The larger models require 80 GB H100. Quantized versions (FP8, INT8) of the larger models are emerging and reduce VRAM requirements by approximately 40 to 50 percent.
Apache 2.0 (Wan 2.1, Mochi-1, Allegro, EasyAnimate) is the most permissive. MIT (Open-Sora 2.0, Pyramid Flow) is similar. HunyuanVideo uses the Tencent Hunyuan Community Licence which permits commercial use but has scale restrictions. LTX-Video uses OpenRAIL-M which has use-case restrictions on harmful content.

Track Your AI Visibility

See how your brand appears across ChatGPT, Claude, Perplexity, and other AI platforms. Start monitoring today.