MiniMax H3 and LTX 2.5 are unlimited on Sogni, Krea 2 learns to edit, the Dark Beasts arrive, and Season 11 opens with 8M SOGNI.
/Sync
VOLUME 22
July / August 2026 Edition
MiniMax H3 Is Unlimited on Sogni
The strongest open-weight video model on the network, included in your plan. No credit counter.
Unmute for sound.
This is the single biggest thing we can offer you right now. MiniMax H3, in both its standard and Turbo builds, runs on Sogni Unlimited fair use. So does LTX 2.5. Two frontier video engines, native audio, no per render math.
Unlimited fair use access to 200+ open-weight models for image, video and music, including H3, LTX 2.5, Krea 2, Z-Image, Flux, Qwen and WAN.
Every plan starts with a 3-day free trial, from $20 a month, across web, iOS / macOS and Android.
Queue up to 128 media at a time, or 512 on Unlimited Pro. Built for batch work, not single renders.
Fair Use, In The Open
You can now see exactly where you stand
Daily and monthly capacity, reset times, and the retail value your plan has already covered.
Fair use limits for Unlimited and Unlimited Pro are no longer a black box. The new Usage page shows your daily and monthly usage, your remaining capacity, and when everything resets.
If you ever reach a limit:
Fast rendering pauses until your limit resets
Relaxed rendering stays unlimited
Premium Spark and SOGNI are unaffected
These limits currently touch only the top 10% of subscribers, and they exist to keep rendering fast for everyone else. Nothing about your plan changes, and there is nothing you need to do.
The open-weight world moved faster than our announcements did. Every model below is live on app.sogni.ai and the API right now, with the plan it runs on marked on each card.
Included in Unlimited
MiniMax H3 and MiniMax H3 Turbo, our smartest video engine, now complete in both builds. Standard H3 gives you the highest quality and the best prompt understanding, while H3 Turbo trades a little of that for speed you can actually iterate on. Turbo now covers text to video, image to video and reference to video, all with native audio. Both builds are covered by your Unlimited plan.
Text to video with native audio, straight out of MiniMax H3. Unmute for sound.
Wan 3 and Wan 3 Enhanced, Alibaba's newest video line, now live on Sogni. Both render 2 to 30 seconds at a fixed 30 fps, up to 1080p, in any aspect ratio from 16:9 to 9:16, with native audio on by default. Both cover text to video, first-frame and first-and-last-frame animation, loose multimodal references and audio-driven generation, taking up to 10 images, 5 videos and 5 audio clips as conditioning.
Three modes on Wan 3 Enhanced: text to video, image to video, reference to video.
The difference is the content policy. Wan 3 Enhanced is the mature-theme build, made by the Wan 3 Enhanced Team for adult creators working on lawful material that standard Wan 3 turns down. It is our premiere mature-theme video model, it needs the Sensitive Content Filter switched off, and it is not restriction-free: everyone depicted must be 18 or over and our Terms of Use still apply.
Both are Premium Spark vendor models, so they run pay as you go rather than on your subscription. Unlimited still helps: 5% off on Unlimited and 10% off on Unlimited Pro, plus a launch credit of $10 for Unlimited and $25 for Unlimited Pro.
LTX 2.5. The standout is multi-shot prompting, which lets you render several hard-cut scenes with continuity from a single prompt.
Substantially sharper, less melty output with better motion
Much better handling of complex prompts
Better quality per compute, plus auto-duration
Full workflow support: t2v, i2v, a2v, v2v, pose and canny controlnet, inpaint, outpaint, LoRAs and more
Multi-shot prompting: several hard-cut scenes with continuity, from one prompt. Unmute for sound.
Next to MiniMax H3: H3 is the smarter, more powerful engine. When LTX 2.5 lands, though, it lands cleaner, with more natural nuance, and it is a lot faster, still at 4K and 60fps. Pricing sits about 20% above LTX 2.3.
NVIDIA RTX VSR Upscaling, our first true pixel upscaler. No reimagining and no invented detail, just clean natural recovery of what is already there. It is roughly 20 times faster than our other upscalers and now goes up to 16K, or 236 MP, which is larger than a 4 by 4 foot print at 300dpi.
832 by 1216 up to 3328 by 4864, nearly 5K on the long edge, in ten seconds of work.
MiniMax Music 3, the new crowned king of open-weight music generation and probably the most production-ready open model right now, at roughly twice the parameters of Ace-Step XL. It sits somewhere between Ace-Step XL 1.5 and Suno. Vocals and production values in open music generation have never sounded this good.
One Obsession v22, a stunning illustration model released only a month ago. It has an obsessive mastery of the 2.5D semi-realistic 3D anime render style, with sublime use of color, contrast, volumetric shading and glossy specular highlights, especially on skin, eyes and hair.
Pay as you go · Not in Unlimited
Seedance 2.5, ByteDance's latest, and a beast for storytelling. One-take creation up to 30 seconds in a single pass with real narrative structure, extendable to multi-minute content. Reference up to 30 images, 10 video clips and 10 audio clips at once, keeping looks and voices consistent, plus timestamp-level editing, green-screen replacement and camera perspective adjustments. As a frontier partner model it runs on Spark Packs rather than your subscription, and Unlimited plans get a significant discount on it.
A skincare spot, one take.
UGC at night, one take.
Workflow Spotlight
MiniMax H3 R2V Turbo: hand it a reference, watch it recast the shot
Reference to video, in two takes, and every run is covered by Unlimited.
MiniMax H3 R2V Turbo, real phone footage plus one reference image, motion untouched.
MiniMax H3 R2V Turbo, a Krea 2 still animated, then recast with a new face on top.
Reference to video is the piece most models still get wrong. Hand MiniMax H3 R2V Turbo a clip and a reference image and it keeps the original performance, the walk, the gesture, the beat, while replacing who is in the frame. The same trick works on generated footage: render with image to video, then recast the character in a second pass.
With the Turbo build live, that loop is finally fast enough to iterate on. And because H3 is an open-weight model on the Supernet, every one of these runs is covered by Sogni Unlimited. That combination is the best reason to be on a plan right now.
Image EditingIncluded in Unlimited
Krea 2 learned to edit, and it kept the face
Krea 2 Identity Edit v1.2, live on Sogni less than 24 hours after it was published.
One input, two context images, three prompts. Outfit swapped, coat layered back on, pose matched to a drawing.
There is an exciting community project turning the fast, powerful Krea 2 Turbo image model into a proper image edit model. Early releases were promising, and v1.2 is the one that delivers. The reviews say so too.
The hardest part of using it is normally getting it running in ComfyUI with new custom nodes. Sogni has your back, and we are live with it in two flavors:
Regular Krea 2 for cleaner, more predictable likeness preservation: product edits, try-on, clothes swap, face swap, scene swap, character sheets.
Dark Beast Krea 2 for the uncensored base: edgier, horror, surreal, adult-oriented or less constrained edits. More stylized, less predictable.
Two uncensored fine-tunes trained by CivitAI creator AiMetatron.
As the creator puts it, these are made “for those who won’t settle for anything less than complete, uncensored reality. Every pixel, every shadow, every nuance is rendered with obsessive precision, creating a world that feels alive, tangible, and dangerously real.”
Dark Beast Krea 2
A 10 to 20 step fine-tune of Krea 2, itself only a month old. It calls itself uncensored, unfiltered and unapologetically hardcore, but it is also a seriously good creative model, making more complex and dynamic art than Krea 2 Turbo itself.
A fine-tune of the very powerful Z-Image Turbo base, adding room for the weird and wonderful to an already diverse model. Its text handling and prompt understanding are next level, and it holds a long prompt better than almost anything else we host.
These two are aggressively uncensored, so unchecking the Mature Content filter is required to use them. Both Krea 2 Identity Edit flavors are uncensored as well, so user discretion is advised.
New LoRAsIncluded in Unlimited
The Krea 2 LoRA library is now 25 strong
Two new community styles, a slider that controls how far the model will go, and our first video LoRA for MiniMax.
Krea 2 is the model our community keeps building on, and the LoRA shelf on Sogni has grown to 25 community fine-tunes. Two of them are brand new this month, and one of them is not a style at all but a control.
New this month
BloomGirls UltraRealism and Purple Grainy Photography
Same portrait, four passes: no LoRA, BloomGirls, Purple Grainy, then both stacked together.
BloomGirls UltraRealism brings smoother skin, brighter bloom, and a polished, aspirational aesthetic.
Purple Grainy Photography adds visible film grain, cooler tones, and a more editorial, analog feel.
Use them individually or stack them together for AI influencer, fashion, beauty and lifestyle content that reads as both realistic and aspirational. They join 23 community LoRAs already on the shelf, all of which run on Krea 2 Turbo.
FilterBypass: turn the safety hedging down like a volume knob
Identical prompt, identical seed. Only the strength value moves: Off, +0.5, +1, +2, +100.
Ask most image models for something and you get a polite version of it. The FilterBypass LoRAs for Krea 2, 2vector and 3vector, make that politeness a value you set. Off is the stock model, +1 commits, +100 holds nothing back.
Not a shock-value gimmick, just a model that renders what you asked for. Horror that lands, emotion that reads, edge that stays sharp.
Uncensored models, unlimited generations, no credit counter watching you work.
First video LoRA
Realism People, for MiniMax video
LoRAs are not just an image thing anymore. Right now there is exactly one video LoRA for MiniMax on Sogni, Realism People, and it pushes toward a more natural, cinematic look: realistic textures, lighting, skin detail and movement. Same prompt, same seed, just less synthetic. More MiniMax video LoRAs land in the next few days.
Set the strength, hit add. Recommended range is +0.6 to +1.
Creative Agent
Stop prompting one image at a time
One brief to Sogni Creative Agent, a whole series on disk, same style in every frame.
The actual output folder: campaign subjects, typographic spots and first frames, all named and numbered.
The Sogni Creative Agent is an installable skill that gives any AI agent production-grade image, video and music generation on the Supernet. You describe the batch in plain English. It plans the run, estimates the cost up front, generates everything, and saves each result as a real named file on your disk, not a preview trapped in a chat window.
One command installs it:
Setup and install sogni.ai/skill
It runs in Claude Code, Claude Desktop and Cowork, Codex CLI, OpenClaw and any SKILL.md aware runtime, with personas and memories that follow you between them. Not a terminal person? In Cowork it shows up as the sogni-creative-agent connector, one toggle away.
The whole brief: count, one shared style, a reference image, a destination folder, aspect ratio, model.
How to prompt for a series
Tell it how many images, that they share one style, and where to save them. Attach a reference image to lock the look. It plans the batch, renders it, and leaves named files on your disk.
One more prompt animates the whole folder. Working at this volume is exactly what Sogni Unlimited is built for.
This is not a 3D video. It is a website reacting to your scroll.
A GPU through a psychedelic tunnel, built by a Sogni Agent, wrapped around our economics essay.
The camera accelerates with the scroll, then hands off to the article. Sped up for this clip.
Most unlimited AI plans quietly disappear, and the reason is structural. On centralized GPU infrastructure every generation costs the provider more, so the math eventually breaks. Powering an Unlimited AI Economy That Works is our case for why the Sogni Supernet is different: independent GPU operators running open models, earning up to 51% of network revenue, on a network that cleared 158 million renders in its first public year. That structure is what makes unlimited hold up instead of quietly getting withdrawn.
The presentation makes the point too. That GPU tunnel is not a video, it is the page reacting to your scroll, and the whole effect was built by your own AI agent working with the Sogni Creative Agent. Give it a link or a brief and it comes back with the scroll-driven page and an MP4 demo.
Season 10 has closed, its airdrop lands September 7, and workers take a bigger cut this time.
Congratulations to everyone who ranked in Season 10. Its airdrop lands at dashboard.sogni.ai on Monday, September 7. Worker revenue share follows the usual 6-week schedule, while unlimited sub earnings pay out on fiscal month close plus 45 days.
Season 11 is now live with the same terms and one change: worker revenue share is up 10%, ambassador share down 10%. August was a breakout month for adoption, driven by Unlimited, and that momentum carries straight in.
The agent revolution is here, and you are its creative director
A packed house, four speakers, and one very persuasive demo.
Mandala Club Singapore: a brief goes in, images, video and finished creative come out.
Mandala Club Singapore: speakers on, wordmark up, and the room got the whole show.
Everyone asks what AI agents can actually do. On Tuesday at Mandala Club Singapore, we showed them.
It was a packed house, with Sogni CEO Mauvis Ledford presenting alongside three other phenomenal speakers. His talk, Give Your AI Agent a Creative Department, walked the room through an agent that runs exactly that. Brief in, storyboard approved, finished video out.
The feedback was great, and the best part was watching people leave with new ideas about what agents can do beyond writing and coding. If you want to build the same thing yourself, it starts with the Creative Agent skill.
