1
LTX-2.5
Lightricks · Video: local text- and image-to-video with sound, up to 4K
$
LTX-2.5 is the most practical open video model for most creators: it makes video and synchronised audio in one pass, goes up to 4K and 20 seconds, and is free for commercial use under $10M annual revenue. Day-one ComfyUI support and a growing set of official control LoRAs make it easy to build on. Picture quality trails the closed leaders (#25 on Artificial Analysis text-to-video), so it is the dependable workhorse rather than the prettiest result.
- Cheapest way in
- Free weights; free commercial use under $10M annual revenue
- Cost per output
- LTX API about $0.13/s at 1080p (Fast) or $0.17/s (Pro), so about $1.30-1.70 for a 10s 1080p clip
- Licence
- LTX-2.x Community License (commercial use under $10M revenue)
- Min VRAM
- 24GB (RTX 3090/4090) with int8; less with heavy offloading
- Max resolution
- 4K (up to 10s); 1080p for longer clips
- Max length
- 20 seconds
- ComfyUI
- Native, with official workflow templates
- Native audio
- Yes
Good at
- Video and audio generated together in one pass
- Up to 4K output and clips up to 20 seconds
- Official ComfyUI workflows plus control LoRAs (restore, upscale, day-to-night, clean plate)
- Free commercial use under $10M annual revenue
Watch out for
- Image quality well behind closed models in blind tests
- Realistically needs a 24GB card; smaller GPUs rely on slow offloading
- Clips over 10 seconds are limited to 1080p
Benchmarks Artificial Analysis Text to Video Arena (with audio): LTX-2.5 Fast #25, Elo 947; LTX-2.5 Pro #26, Elo 944 (#2 open-weights model after MiniMax H3) · Artificial Analysis Image to Video Arena: LTX-2.5 Fast #25, Elo 918; LTX-2.5 Pro #26, Elo 910 · Arena Text to Video: LTX-2 19B (older version) #41, 1154 +/-8 (page dated 21 Sep 2026); LTX-2.5 not yet listed
Prices artificialanalysis.ai, checked 1 November 2026
LTX-2.5 website Work made with it on AI LIVE
2
FLUX 3 Image Klein
Black Forest Labs · Image: fast local 4K images and multi-reference edits on 8GB GPUs
$
Klein is the highest-ranked open-weights model on the Artificial Analysis text-to-image arena (Elo 1043) and also scores well for editing (Elo 1072). Released on 15 October 2026, it generates natively at 4K, takes up to 10 reference images and, in its INT8 build, runs in ComfyUI on cards with as little as 8GB of VRAM. It is free for commercial output for businesses under $5M revenue.
- Cheapest way in
- Free weights; commercial output free under $5M annual revenue
- Cost per output
- BFL API about $0.009 per image
- Licence
- FLUX Kommunity License v1.0
- Min VRAM
- 8GB (INT8 in ComfyUI); about 25-31GB at full precision
- Max resolution
- Native 4K
- ComfyUI
- Native since 15 Oct 2026, with official templates
- References
- Up to 10 images
- Parameters
- 5.6B
Good at
- Top open-weights score on Artificial Analysis text-to-image
- Native 4K with up to 10 reference images
- Runs on 8GB GPUs in ComfyUI (INT8 build)
- Base checkpoint for LoRA training, backed by the biggest open image ecosystem
Watch out for
- Still well behind the closed leaders (#19-26 on text-to-image)
- Paid licence needed above $5M revenue
- Commercial fine-tunes and redistribution need separate permission
Benchmarks Artificial Analysis Text to Image Arena: #21 (range 19-26), Elo 1043; top open-weights model · Artificial Analysis Image Editing Arena: #27 (range 24-32), Elo 1072
Prices artificialanalysis.ai, checked 1 November 2026
FLUX 3 Image Klein website
3
MiniMax H3
MiniMax · Video: the best-looking open video model, where its licence allows
$
H3 is the strongest open-weights video model by a wide margin: #1 on Arena image-to-video and #3 on Artificial Analysis text-to-video, ahead of most closed models. The open release makes 768p clips of 4-15 seconds with stereo dialogue and sound, and runs in ComfyUI. Its licence excludes the US, EU, UK and South Korea unless MiniMax grants permission, which rules it out for many creators.
- Cheapest way in
- Free weights; community licence, not valid in the US, EU, UK or Korea
- Cost per output
- MiniMax API about $0.08/s at 768p and $0.13/s at 2K, so about $1.20 for a 15s 768p clip
- Licence
- MiniMax H3 Community License (excludes US, EU, UK, South Korea)
- Min VRAM
- High-end GPU: INT8 model is 32GB (19.5GB pruned) plus text encoder
- Max resolution
- 768p (open weights)
- Max length
- 15 seconds at 24fps
- ComfyUI
- Native since 3 Aug 2026
- Native audio
- Yes, 32kHz stereo with dialogue
Good at
- Top-tier quality, close to the best closed models
- Native stereo audio with dialogue in 11 languages
- Text, first/last-frame and reference-to-video checkpoints
- Native ComfyUI support from release
Watch out for
- Licence does not cover the US, EU, UK or South Korea without approval
- Heavy: the INT8 model file alone is about 32GB (19.5GB pruned)
- Open weights stop at 768p; 2K is API only
Benchmarks Arena Image to Video: #1 overall, 1495 +/-5 (page dated 21 Sep 2026) · Arena Text to Video: #8 overall, 1460 +/-9; top open model (page dated 21 Sep 2026) · Artificial Analysis Text to Video Arena (with audio): #3 overall, Elo 1161 (768p); top open-weights model · Artificial Analysis Image to Video Arena: #4 overall, Elo 1128 (768p)
Prices platform.minimax.io, checked 1 November 2026
MiniMax H3 website Work made with it on AI LIVE
4
Wan 2.2
Alibaba · Video: fully free Apache-licensed video, LoRAs and character animation
$
Wan 2.2 is still the newest open Wan model, since Wan 2.5 to 3.0 are API only. Its Apache 2.0 licence allows commercial use with no revenue cap, and it has the largest library of community LoRAs and ComfyUI workflows of any open video model. It scores below H3 and LTX in blind tests (#42 on Arena text-to-video) and makes silent 5-second clips at up to 720p.
- Cheapest way in
- Free weights; Apache 2.0, unrestricted commercial use
- Cost per output
- fal about $0.08/s at 720p or $0.04/s at 480p, so about $0.40 for a 5s 720p clip
- Licence
- Apache 2.0
- Min VRAM
- 8GB for the 5B model with ComfyUI offloading; much more for 14B
- Max resolution
- 720p
- Max length
- 5 seconds
- ComfyUI
- Native, with official templates
- Native audio
- No
Good at
- Apache 2.0: commercial use with no revenue cap
- Biggest ecosystem of LoRAs, fine-tunes and ComfyUI workflows
- 5B model fits 8GB GPUs with ComfyUI offloading
- Wan-Animate-2 (August 2026) adds character animation from a driving video
Watch out for
- No native audio
- 5-second clips at up to 720p
- Clearly behind 2026 models in blind tests
Benchmarks Arena Text to Video: Wan 2.2 A14B #42, 1132 +/-15 (page dated 21 Sep 2026) · Arena Image to Video: Wan 2.2 A14B #43, 1170 +/-10 (page dated 21 Sep 2026)
Prices fal.ai, checked 1 November 2026
Wan 2.2 website
5
Qwen-Image-Edit-2511
Alibaba · Image: prompt-based photo editing you can run locally
$
Qwen's open image models are the strongest free route to prompt-based editing: Qwen-Image-2.1 is the top open model on both Arena image edit (#20) and Arena text-to-image (#19). That 2.1 release is research-only, so for paid work the Apache 2.0 Qwen-Image-Edit-2511 and Qwen-Image 2512 are the safe choices (#41 and #45 on Arena). All run in ComfyUI, though the 20B models want a 16-24GB card.
- Cheapest way in
- Free weights; Edit-2511 Apache 2.0, Qwen-Image-2.1 research use only
- Cost per output
- fal $0.03 per megapixel for Edit-2511, about 3 cents per 1MP edit
- Licence
- Apache 2.0 (Edit-2511, 2512); research only (2.1)
- Min VRAM
- 12-16GB with GGUF builds; 24GB for FP8
- Max resolution
- 2048x2048 (2.1)
- ComfyUI
- Native
- Parameters
- 20B (Edit-2511)
- Editing
- Instruction-based, with multi-image input
Good at
- Highest open-model scores for editing on Arena
- Edit-2511 and Qwen-Image 2512 are Apache 2.0
- Multi-image input for edits
- Native 2K output with transparency (2.1)
Watch out for
- The strongest version (2.1) is non-commercial
- Commercial versions trail the closed leaders by a wide margin
- Large 20B models are slow on mid-range GPUs
Benchmarks Arena Image Edit (single image): Qwen-Image-2.1 #20, 1362 +/-6 (preliminary); Qwen-Image-Edit 2511 #41, 1233 +/-3 (page dated 23 Oct 2026) · Arena Text to Image: Qwen-Image-2.1 #19, 1225 +/-6 (preliminary); Qwen-Image 2512 #45, 1125 +/-3 (page dated 26 Oct 2026) · Artificial Analysis Image Editing Arena: Qwen-Image-2.1 #24, Elo 1081; Qwen Image Edit Plus 2511 #52, Elo 991
Prices fal.ai, checked 1 November 2026
Qwen-Image-Edit-2511 website
6
Z-Image Turbo
Alibaba · Image: very fast, low-cost generation on mid-range GPUs
$
Z-Image Turbo is a 6B model that makes an image in 8 steps and fits comfortably on a 16GB card, which makes it one of the quickest decent open models to run at home. It is Apache 2.0, so commercial use is unrestricted, and it renders English and Chinese text. It sits mid-table in blind tests (#59 on Arena text-to-image), trading top-end quality for speed.
- Cheapest way in
- Free weights; Apache 2.0, unrestricted commercial use
- Cost per output
- fal $0.005 per megapixel, about half a cent per 1MP image
- Licence
- Apache 2.0
- Min VRAM
- 16GB; about 6GB with quantised builds
- ComfyUI
- Native, with official workflows
- Parameters
- 6B
- Steps
- 8
- Text rendering
- English and Chinese
Good at
- Apache 2.0 licence
- 8-step generation
- Fits 16GB cards; quantised builds go lower
- English and Chinese text rendering
Watch out for
- Mid-table quality in blind tests
- No open edit model released yet
- Released November 2025, older than most rivals here
Benchmarks Arena Text to Image: #59, 1082 +/-5 (page dated 26 Oct 2026) · Artificial Analysis Text to Image Arena: Rank range 70-76, Elo 936
Prices fal.ai, checked 1 November 2026
Z-Image Turbo website
7
Fish Audio Drama 3
Fish Audio · Audio: expressive multi-character dialogue and voice cloning
$$
Drama 3 is the highest-ranked open-weights voice model on the Artificial Analysis speech arena (#6, Elo 1278), level with many paid services. It voices up to 16 characters in one scene from bracketed acting directions across 83 languages, and a community ComfyUI node runs it on 8GB cards by keeping its experts in system RAM. The licence is free for commercial use under $500,000 annual revenue.
- Cheapest way in
- Free weights; commercial use free under $500k annual revenue
- Cost per output
- Fish Audio API $10 per 1M UTF-8 bytes until 12 Nov 2026, then $50 (about $0.05 per 1,000 characters)
- Licence
- Fish Audio Community License (commercial use under $500k revenue)
- Min VRAM
- 8GB with ComfyUI-Drama3 (experts held in system RAM)
- ComfyUI
- Community node (ComfyUI-Drama3)
- Languages
- 83
- Multiple speakers
- Up to 16 characters
- Parameters
- 35B mixture-of-experts, about 3B active
Good at
- Best open voice model in blind tests
- Up to 16 characters per scene with inline acting direction
- Voice cloning from reference audio
- Runs on 8GB GPUs via the ComfyUI-Drama3 node
Watch out for
- Commercial licence needed above $500,000 revenue or to resell as a voice API
- Official package is about 40GB on disk
- ComfyUI node is community-made, not official
Benchmarks Artificial Analysis Provider Voice Arena (Text to Speech): #6, Elo 1278; top open-weights model
Prices docs.fish.audio, checked 1 November 2026
Fish Audio Drama 3 website Work made with it on AI LIVE
8
SeedVR2
ByteDance Seed · Video: one-step upscaling and restoration for clips and stills
$
SeedVR2 is a one-step diffusion restorer that upscales video while keeping frames consistent, and works on still images too. It is Apache 2.0, comes in 3B and 7B sizes and has native ComfyUI nodes, which makes it a common final pass after LTX or Wan. It can over-sharpen clean AI footage, so it suits soft or low-resolution sources best.
- Cheapest way in
- Free weights; Apache 2.0, unrestricted commercial use
- Cost per output
- fal $0.001 per megapixel of video, about $0.25 to upscale 5 seconds of 1080p
- Licence
- Apache 2.0
- Min VRAM
- 16GB possible with the 3B model; 24GB+ recommended for video
- Sizes
- 3B and 7B
- ComfyUI
- Native nodes
- Input
- Video and still images
Good at
- Apache 2.0 licence
- One step, so much faster than multi-step upscalers
- Upscales both video and images
- Native ComfyUI nodes with FP8 and INT8 builds
Watch out for
- Can over-sharpen or invent detail on clean 720p AI footage
- Struggles with heavy degradation and very fast motion
- Video batches want 24GB or more for smooth results
Prices fal.ai, checked 1 November 2026
SeedVR2 website
9
MiniMax Music 3.0
MiniMax · Audio: full songs with vocals up to five minutes
$
MiniMax Music 3.0 is the top open-weights model on the Artificial Analysis music arenas, at Elo 1000 (#9 on vocals), behind Suno and other closed leaders. It writes complete songs with vocals up to five minutes, fits under 24GB of VRAM (or 8GB with slower streaming) and runs in ComfyUI. Commercial use is allowed with no regional limits, though commercial apps must show the model name and firms above $20M revenue need a deal.
- Cheapest way in
- Free weights; commercial use allowed under $20M annual revenue
- Cost per output
- Hosted APIs from about $0.20 per song; MiniMax's own $0.15 per song API is closed to new users
- Licence
- MiniMax-Music3 Community License
- Min VRAM
- Under 24GB; 8GB with layer streaming (slower)
- Max length
- 5 minutes
- ComfyUI
- Supported, with tutorials and a Comfy-Org build
- Vocals
- Yes
Good at
- Top open music model in blind tests
- Vocals with control over gender, timbre, harmony and effects
- Fits under 24GB; streams on 8GB cards
- No regional restrictions in the licence
Watch out for
- Clearly behind Suno v6 in blind tests
- MiniMax's own music API is closed to new users
- Commercial apps must display the MiniMax-Music3 name
Benchmarks Artificial Analysis Music Arena: Vocals: #9, Elo 1000 (anchor); top open-weights model · Artificial Analysis Music Arena: Instrumental: #13 (range 8-15), Elo 1000
Prices muapi.ai, checked 1 November 2026
MiniMax Music 3.0 website
10
ACE-Step 1.5 XL
ACE Studio and StepFun · Audio: MIT-licensed music up to 10 minutes on modest GPUs
$
ACE-Step 1.5 is the most permissive open music model: MIT-licensed, with outputs cleared for commercial use and training data its makers describe as licensed, royalty-free or public domain. It makes 10 seconds to 10 minutes of music with lyrics in 50+ languages, in under 10 seconds on an RTX 3090, and the small models run on very little VRAM. It has no score on the public music arenas yet.
- Cheapest way in
- Free weights; MIT licence, unrestricted commercial use
- Cost per output
- Runware about $0.004 per text-to-music generation
- Licence
- MIT
- Min VRAM
- Under 6GB (2B model); 12-24GB for XL
- Max length
- 10 minutes
- ComfyUI
- Native, with official templates
- Languages
- 50+ for lyrics
Good at
- MIT licence; outputs usable commercially
- Very fast generation
- Songs up to 10 minutes, lyrics in 50+ languages
- Native ComfyUI templates
Watch out for
- No independent blind-test score yet
- The XL model wants 20-24GB for best quality
- Cover mode not yet available in ComfyUI
Prices runware.ai, checked 1 November 2026
ACE-Step 1.5 XL website
