Wan 3.0 is here! 🚀
In short: Alibaba has officially launched the public beta of Wan 3.0. The new model can generate videos with audio up to 30 seconds long in 480p, 720p, and 1080p.
New features also include automatic duration selection, video continuation, and improved consistency of characters, objects, and scenes across longer clips.
https://article.9466.com/news/aKweEQlm
API pricing:
480p — around $0.04 per second
720p — around $0.08 per second
1080p — around $0.17 per second
That means a maximum 30-second video costs approximately $1.25 at 480p, $2.50 at 720p, and around $5 at 1080p.
But there’s a catch: when using a video reference, Alibaba charges not only for the generated output but also for the duration of the uploaded reference video.
The formula is: input video + generated video.
For example, a 10-second reference combined with a 30-second 1080p output would cost approximately $6.70.
In short: Alibaba has officially launched the public beta of Wan 3.0. The new model can generate videos with audio up to 30 seconds long in 480p, 720p, and 1080p.
New features also include automatic duration selection, video continuation, and improved consistency of characters, objects, and scenes across longer clips.
https://article.9466.com/news/aKweEQlm
API pricing:
480p — around $0.04 per second
720p — around $0.08 per second
1080p — around $0.17 per second
That means a maximum 30-second video costs approximately $1.25 at 480p, $2.50 at 720p, and around $5 at 1080p.
But there’s a catch: when using a video reference, Alibaba charges not only for the generated output but also for the duration of the uploaded reference video.
The formula is: input video + generated video.
For example, a 10-second reference combined with a 30-second 1080p output would cost approximately $6.70.
👍665🎉612❤588🔥578
Media is too big
VIEW IN TELEGRAM
LTX fires back with LTX-2.5 — a new open 22B video/audio world model.
Highlights: better motion quality, improved prompt understanding, native multishot with character/voice consistency, 4K HDR, RAW workflow, and precise video editing.
The big flex is speed + local inference: 16 GB VRAM minimum, ComfyUI support, open weights, and 10s I2V in 6.8s on 2× GB200.
LTX also published comparisons vs H3, Seedance, Kling, Veo and Grok — and Reddit is already roasting the methodology.
Marketing or not, this one is worth testing.
Highlights: better motion quality, improved prompt understanding, native multishot with character/voice consistency, 4K HDR, RAW workflow, and precise video editing.
The big flex is speed + local inference: 16 GB VRAM minimum, ComfyUI support, open weights, and 10s I2V in 6.8s on 2× GB200.
LTX also published comparisons vs H3, Seedance, Kling, Veo and Grok — and Reddit is already roasting the methodology.
Marketing or not, this one is worth testing.
❤164🎉163🔥162👍159
Dreamina just added Seedance 2.5 in 1080p 🚀
Using the current exchange rate, the pricing comes out to roughly:
≈ $0.95 per second
So a 30-second video costs around $28.50–$30.90.
In short: a good working estimate for Seedance 2.5 1080p right now is about $0.95/sec, or roughly $29–31 for 30 seconds.
Using the current exchange rate, the pricing comes out to roughly:
≈ $0.95 per second
So a 30-second video costs around $28.50–$30.90.
In short: a good working estimate for Seedance 2.5 1080p right now is about $0.95/sec, or roughly $29–31 for 30 seconds.
👍58❤57🔥51🎉51
Media is too big
VIEW IN TELEGRAM
Hyperion 2.5:
Topaz has released Hyperion 2.5 — an SDR→HDR model aimed at 8-bit AI video. Instead of simply repacking footage into 10-bit, it reconstructs additional luminance and color information and outputs 10-bit ProRes, H.265, or 16-bit EXR.
This matters because heavily compressed 8-bit GenAI footage quickly falls apart during aggressive grading, compositing, or HDR work. Hyperion processes it before grading, making it behave more like a high-bit-depth source.
Darren Aronofsky’s Primordial Soup has already used it to convert GenAI footage to 10-bit ProRes before heavy color work and reduce visible banding.
It can’t recover data that never existed — this is AI reconstruction, essentially a smart HDR remaster, not camera RAW.
There are still some 16-bit EXR color-management issues, but it’s a promising step toward making GenAI footage more suitable for high-end post. Hyperion 2.5 is available in Topaz Video, locally or via cloud rendering.
Topaz has released Hyperion 2.5 — an SDR→HDR model aimed at 8-bit AI video. Instead of simply repacking footage into 10-bit, it reconstructs additional luminance and color information and outputs 10-bit ProRes, H.265, or 16-bit EXR.
This matters because heavily compressed 8-bit GenAI footage quickly falls apart during aggressive grading, compositing, or HDR work. Hyperion processes it before grading, making it behave more like a high-bit-depth source.
Darren Aronofsky’s Primordial Soup has already used it to convert GenAI footage to 10-bit ProRes before heavy color work and reduce visible banding.
It can’t recover data that never existed — this is AI reconstruction, essentially a smart HDR remaster, not camera RAW.
There are still some 16-bit EXR color-management issues, but it’s a promising step toward making GenAI footage more suitable for high-end post. Hyperion 2.5 is available in Topaz Video, locally or via cloud rendering.
🎉63👍59🔥55❤53
Competition is great.
It looks like OpenAI and Anthropic are competing to see who can suck up to users harder, just to keep them from jumping off their train and hopping onto the competitor’s.
OpenAI went ahead and dropped API prices for SOL 5.6 by 20% for a whole THREE months.
Anthropic was like: fine, we’re extending the increased weekly Claude Code limits through August 31.
Users: message received, keep it coming. Ox Alpha is completely free anyway, so we’ll be waiting for even more concessions.
It looks like OpenAI and Anthropic are competing to see who can suck up to users harder, just to keep them from jumping off their train and hopping onto the competitor’s.
OpenAI went ahead and dropped API prices for SOL 5.6 by 20% for a whole THREE months.
Anthropic was like: fine, we’re extending the increased weekly Claude Code limits through August 31.
Users: message received, keep it coming. Ox Alpha is completely free anyway, so we’ll be waiting for even more concessions.
🎉286🔥280👍265❤243
MiniMax H3 Max: 36x faster — and the weights are going open
Regular MiniMax H3 in Minimax Design cloud vs H3 Max on fal.
Same settings: 15 sec, 720p.
H3 took around 6 minutes. H3 Max took around 10 seconds.
That’s roughly 36x faster, with comparable quality to my eye.
Important caveat: this is not a clean model benchmark. These are different production stacks. Still, 36x is almost certainly not just “more GPUs”.
Big news:
fal confirmed H3 Max weights will be released.
Asked whether Ref2V will also be open:
“yes, working on it.”
So we may get an open H3 Max checkpoint, and later Ref2V too.
The key question: how much of that speed is in the model, and how much in fal’s stack?
I wouldn’t expect H3 Max to run 36x faster locally.
The gain likely comes from fewer sampling steps, optimized attention, fused CUDA/Triton kernels, mixed precision, parallelism, VAE and memory optimizations.
Part of the speed is in the weights, part is in fal’s infrastructure.
That ratio will matter most after release.
Regular MiniMax H3 in Minimax Design cloud vs H3 Max on fal.
Same settings: 15 sec, 720p.
H3 took around 6 minutes. H3 Max took around 10 seconds.
That’s roughly 36x faster, with comparable quality to my eye.
Important caveat: this is not a clean model benchmark. These are different production stacks. Still, 36x is almost certainly not just “more GPUs”.
Big news:
fal confirmed H3 Max weights will be released.
Asked whether Ref2V will also be open:
“yes, working on it.”
So we may get an open H3 Max checkpoint, and later Ref2V too.
The key question: how much of that speed is in the model, and how much in fal’s stack?
I wouldn’t expect H3 Max to run 36x faster locally.
The gain likely comes from fewer sampling steps, optimized attention, fused CUDA/Triton kernels, mixed precision, parallelism, VAE and memory optimizations.
Part of the speed is in the weights, part is in fal’s infrastructure.
That ratio will matter most after release.
🎉361👍359❤350🔥319
THIS is WILD.
Not another candle, not another skincare bottle. I am building a full launch kit for a futuristic smart ring inside Lovart with Seedance 2.5.
Lovart is not just a video generator. It is an AI design agent, so I start with one brief: matte graphite smart ring, holographic charging case, neon blue interface, luxury tech launch, 30-second hero video, posters, and social ads.
Try Lovart here
Watch full video https://www.tiktok.com/@web3.world.yt/video/7680504360232701205
Not another candle, not another skincare bottle. I am building a full launch kit for a futuristic smart ring inside Lovart with Seedance 2.5.
Lovart is not just a video generator. It is an AI design agent, so I start with one brief: matte graphite smart ring, holographic charging case, neon blue interface, luxury tech launch, 30-second hero video, posters, and social ads.
Try Lovart here
Watch full video https://www.tiktok.com/@web3.world.yt/video/7680504360232701205
❤125🔥120👍114🎉110
🚨 OpenAI introduced GPT-6 Astra — and this feels like more than just another model upgrade.
Astra is built not only to answer questions, but to actually do work: coding, browsing, research, documents, and complex multi-step tasks.
Some standout numbers: 99.9% on ARC-AGI-3, 97.6% on FrontierMath Tier 4, 100% on ExploitBench, a 1.05M-token context window, and up to 128K output tokens.
It’s also OpenAI’s first model to reach Critical cyber capability, meaning it can potentially discover and exploit previously unknown vulnerabilities with the right tools.
The bigger story is the shift itself: AI is moving from “smart assistant” to something you can increasingly delegate entire workflows to.
Astra is built not only to answer questions, but to actually do work: coding, browsing, research, documents, and complex multi-step tasks.
Some standout numbers: 99.9% on ARC-AGI-3, 97.6% on FrontierMath Tier 4, 100% on ExploitBench, a 1.05M-token context window, and up to 128K output tokens.
It’s also OpenAI’s first model to reach Critical cyber capability, meaning it can potentially discover and exploit previously unknown vulnerabilities with the right tools.
The bigger story is the shift itself: AI is moving from “smart assistant” to something you can increasingly delegate entire workflows to.
🔥212❤202👍198🎉196
GPT-6 Astra: messy rollout, resets, and Plus segregation
Three updates: one okay, one very good, one very bad.
OpenAI has revealed the limits. $200 Pro gets 200 Astra messages/week. The new $100 Pro gets 50/week, shared with GPT-5.6 Sol Pro. Business Standard gets 15/month, Premium 50/week. So Astra isn’t unlimited even on $200 Pro.
Good news: OpenAI says it will compensate rollout delays. For every day Astra is unavailable on a paid account, users get one banked reset to use later
Bad news for Plus: OpenAI’s latest help article says, “It is not included with ChatGPT Plus in Chat.” Astra appears in regular Chat as GPT-6 Pro only on Pro $100, Pro $200, Business, and Enterprise
Meanwhile, the launch page still says Astra is coming to Plus, Pro, Business, and Enterprise. Where Plus gets access is unclear. Maybe Work or Codex?? Current docs confirm Astra there only for Pro
So: Astra limits are known, but Plus users waiting for it in regular ChatGPT appear out of luck for now. Codex access for Plus is unclear too
Three updates: one okay, one very good, one very bad.
OpenAI has revealed the limits. $200 Pro gets 200 Astra messages/week. The new $100 Pro gets 50/week, shared with GPT-5.6 Sol Pro. Business Standard gets 15/month, Premium 50/week. So Astra isn’t unlimited even on $200 Pro.
Good news: OpenAI says it will compensate rollout delays. For every day Astra is unavailable on a paid account, users get one banked reset to use later
Bad news for Plus: OpenAI’s latest help article says, “It is not included with ChatGPT Plus in Chat.” Astra appears in regular Chat as GPT-6 Pro only on Pro $100, Pro $200, Business, and Enterprise
Meanwhile, the launch page still says Astra is coming to Plus, Pro, Business, and Enterprise. Where Plus gets access is unclear. Maybe Work or Codex?? Current docs confirm Astra there only for Pro
So: Astra limits are known, but Plus users waiting for it in regular ChatGPT appear out of luck for now. Codex access for Plus is unclear too
❤362🎉347👍343🔥332
This media is not supported in your browser
VIEW IN TELEGRAM
So, how are the modelers doing with the new horsepower?
Or rather, with the new Astra.
Blender MCP + Astra + a bit of Computer Use: feed it a sketch of a steam locomotive, and a few minutes later you get 3,295 objects.
https://x.com/tomkrcha/status/2095756085890310311
I think that’s it… we’re done.
Or rather, with the new Astra.
Blender MCP + Astra + a bit of Computer Use: feed it a sketch of a steam locomotive, and a few minutes later you get 3,295 objects.
https://x.com/tomkrcha/status/2095756085890310311
I think that’s it… we’re done.
❤127👍125🎉117🔥110
This media is not supported in your browser
VIEW IN TELEGRAM
Viggle is back with an open model — Viggle-Animate — and it’s pretty damn capable.
For the first time, they’re releasing weights + code.
It replaces characters in video using MiniMax H3. Take a video, grab one frame, manually redraw the character as anything you want, and the model propagates that replacement through the entire clip while preserving motion, camera, and timing.
No pose skeleton, segmentation, face tracking, depth, or text prompt. Just the video + one modified frame.
Under the hood: a 33.1B MiniMax H3 ref2va fine-tune + DMD2 LoRA. Distillation brings it down to 4 sampling steps.
It’s 6× faster than Wan2.2-Animate-14B.
The catch: hardware. BF16 needs ~62 GiB for the transformer and peaks at ~80.1 GiB VRAM. They recommend 96 GB+ or CPU offload. It reportedly runs on a 32 GB RTX 5090 via NVFP4, but that quant isn’t public yet.
Also, you can’t use a normal character portrait. The reference has to be a frame from the original video, already redrawn as the target character.
https://viggle.ai/h3
For the first time, they’re releasing weights + code.
It replaces characters in video using MiniMax H3. Take a video, grab one frame, manually redraw the character as anything you want, and the model propagates that replacement through the entire clip while preserving motion, camera, and timing.
No pose skeleton, segmentation, face tracking, depth, or text prompt. Just the video + one modified frame.
Under the hood: a 33.1B MiniMax H3 ref2va fine-tune + DMD2 LoRA. Distillation brings it down to 4 sampling steps.
It’s 6× faster than Wan2.2-Animate-14B.
The catch: hardware. BF16 needs ~62 GiB for the transformer and peaks at ~80.1 GiB VRAM. They recommend 96 GB+ or CPU offload. It reportedly runs on a 32 GB RTX 5090 via NVFP4, but that quant isn’t public yet.
Also, you can’t use a normal character portrait. The reference has to be a frame from the original video, already redrawn as the target character.
https://viggle.ai/h3
❤57👍57🔥50🎉46
This media is not supported in your browser
VIEW IN TELEGRAM
FLUX.3 Video Edit model
It can take an existing video and, based on a prompt, change objects, characters, backgrounds, materials, text, effects, and even events within the scene. You can also change dialogue and translate conversations with lip sync.
The biggest bombshell: $0.03 per second. Dirt cheap!
So:
5 sec — $0.15
10 sec — $0.30
15 sec — $0.45
But.
It currently works with clips up to 15 seconds, and input videos above 720p are downscaled before processing. No masks, additional references, or resolution selection yet.
In short:
max. 15 sec
max. 50 MiB
up to 720p in the edit pipeline
24 fps
no seed control
apparently no output resolution or duration control (need to dig into this)
no masks
no additional reference images
no video-only motion/style reference
not an extension model — continuing a clip requires FLUX 3 Video Continuation
You can run the result through FLUX Video Upscale separately, up to 4K — but that’s extra $$$.
It can take an existing video and, based on a prompt, change objects, characters, backgrounds, materials, text, effects, and even events within the scene. You can also change dialogue and translate conversations with lip sync.
The biggest bombshell: $0.03 per second. Dirt cheap!
So:
5 sec — $0.15
10 sec — $0.30
15 sec — $0.45
But.
It currently works with clips up to 15 seconds, and input videos above 720p are downscaled before processing. No masks, additional references, or resolution selection yet.
In short:
max. 15 sec
max. 50 MiB
up to 720p in the edit pipeline
24 fps
no seed control
apparently no output resolution or duration control (need to dig into this)
no masks
no additional reference images
no video-only motion/style reference
not an extension model — continuing a clip requires FLUX 3 Video Continuation
You can run the result through FLUX Video Upscale separately, up to 4K — but that’s extra $$$.
❤167🔥166🎉166👍138
Media is too big
VIEW IN TELEGRAM
Google just dropped Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking.
Gemini 3.8 Live is fast voice-to-voice, understands video/images almost in real time, and automatically switches between 97 languages.
(Siri still has ONE.)
It can call tools and APIs in the background without interrupting the conversation.
Extended Thinking is the version for complex tasks: it can reason and speak simultaneously, handle multi-step processes, and verbally comment on its progress.
There’s a Gemini API and AI Studio.
Extended Thinking is also coming to Gemini Live, Docs, Gmail, and Keep.
API pricing is the same for 3.8 Live and Extended Thinking:
Input audio — $0.005/min
Output audio — $0.018/min
Video/image input — around $0.002/min
Text — $0.75 per 1M input tokens and $4.50 per 1M output tokens, including thinking tokens.
Google wants to turn Gemini Live from a “voice chat” into a full-fledged realtime agent that can see, think, talk, and do things in parallel through tools.
Gemini 3.8 Live is fast voice-to-voice, understands video/images almost in real time, and automatically switches between 97 languages.
(Siri still has ONE.)
It can call tools and APIs in the background without interrupting the conversation.
Extended Thinking is the version for complex tasks: it can reason and speak simultaneously, handle multi-step processes, and verbally comment on its progress.
There’s a Gemini API and AI Studio.
Extended Thinking is also coming to Gemini Live, Docs, Gmail, and Keep.
API pricing is the same for 3.8 Live and Extended Thinking:
Input audio — $0.005/min
Output audio — $0.018/min
Video/image input — around $0.002/min
Text — $0.75 per 1M input tokens and $4.50 per 1M output tokens, including thinking tokens.
Google wants to turn Gemini Live from a “voice chat” into a full-fledged realtime agent that can see, think, talk, and do things in parallel through tools.
❤90👍87🎉86🔥71
This media is not supported in your browser
VIEW IN TELEGRAM
Buttery-smooth 120 FPS Three.js arena action, running entirely in your browser. Meet Arena Shooter.
4.6 BILLION Astra tokens burned.
20+ vehicles to capture, zombies, day/night combat, 20+ upgrades & perks, and LLM bots trying to destroy you.
No signup. No install. Just open and fight.
Try https://vortex.channel/vr/soloplay/
4.6 BILLION Astra tokens burned.
20+ vehicles to capture, zombies, day/night combat, 20+ upgrades & perks, and LLM bots trying to destroy you.
No signup. No install. Just open and fight.
Try https://vortex.channel/vr/soloplay/
🎉197❤195👍190🔥186
🚨 Qwen-Image 2.1
Alibaba just dropped Qwen-Image 2.1 — a unified text-to-image + image editing model with only 7B parameters.
🔥 Native 2K
🔥 Up to 10 reference images
🔥 Better identity & consistency
🔥 Masks/scribbles for local editing
🔥 Native RGBA + transparency
🔥 ComfyUI support from day one
The really interesting part: RGBA editing and object extraction from regular RGB images. Need to test this one.
💾 VRAM-wise, 12 GB looks like a realistic target with quantization/offloading. 16 GB should be very comfortable.
The previous Qwen Image was ~20B. This one is just 7B.
🔗 https://modelscope.ai/models/Qwen/Qwen-Image-2.1/summary
Alibaba just dropped Qwen-Image 2.1 — a unified text-to-image + image editing model with only 7B parameters.
🔥 Native 2K
🔥 Up to 10 reference images
🔥 Better identity & consistency
🔥 Masks/scribbles for local editing
🔥 Native RGBA + transparency
🔥 ComfyUI support from day one
The really interesting part: RGBA editing and object extraction from regular RGB images. Need to test this one.
💾 VRAM-wise, 12 GB looks like a realistic target with quantization/offloading. 16 GB should be very comfortable.
The previous Qwen Image was ~20B. This one is just 7B.
🔗 https://modelscope.ai/models/Qwen/Qwen-Image-2.1/summary
🔥569👍565❤521🎉516