Gemini 4 Carbon, Alpha Genome 2, realtime voices, new TTS, ACE Music, Seedream 5.5: AI NEWS
Test Gemini 4 Carbon against your actual workflow before paying; this review found strong multimodal skills but uneven results on long creative tasks.
- Use Gemini 4 Hydrogen for faster everyday tasks; the video says it is available in the Gemini app and Google Search, including the free plan.
- Try open voice models Haru 4B or Sofon Real Time for interruptible speech; Zonos 3 Mini is a lightweight option for text-to-speech.
- For local experiments, check the hardware first: models discussed range from lightweight voice tools to 64–86 GB image and world models.
The review offers hands-on comparisons, but the many brief news summaries are covered by these notes. · Best for: AI practitioners comparing models and tools for coding, media, voice, research, and robotics.
A roundup of new AI models and tools across 3D, video, image generation, voice, music, robotics, and science. The creator’s Gemini 4 Carbon tests were mixed: it handled image analysis well, but struggled with several long creative and computer-use tasks.
3D, video upscaling, and Gemini 4 Carbon tests1:05
- Budget 3D allocates less compute to unimportant regions; the presenter says its tests showed up to about 2× faster generation and lower GPU memory use. Two versions are described: a quality-focused version and Turbo with five steps.
- Reshoot 4D reconstructs footage in 3D, then lets users set a virtual camera path and adjust focus, aperture, and lens effects. Splat Anything reconstructs scenes from multiple photos as Gaussian splats.
- Open H3 Regenerate 2K IC is presented as better at retaining faces and fine details when upscaling video.
- In the presenter’s tests, Gemini 4 Carbon struggled with an explainer animation, browser piano playing, a music project’s vocals, and a procedural Unreal Engine game. It did well on an image-based frog-finding task but poorly on a brain-scan classification test.
- The presenter says Carbon is among the leading frontier models, but found high-mode Antigravity runs slow; several tests took over an hour. The video reports that access required Google’s Ultra plan at the time.
Image generation and faster Gemini model16:46
- The Seedream 5.5 segment describes editing specific image details while keeping the rest consistent, plus transparent backgrounds and separation into up to 100 layers. This segment is a sponsor read.
- Gemini 4 Hydrogen is described as a smaller, faster model for everyday tasks, with text, image, audio, and video input and a context window of up to one million tokens. The video says it was available through the Gemini app and Google Search, including the free plan.
Realtime voice and text-to-speech19:41
- Haru 4B processes voice and generates speech directly, enabling simultaneous listening and speaking; the video says it can run locally with about 12 GB of VRAM and has an Apple silicon implementation.
- Fish Audio Live 3 supports interruptible conversation and acknowledgements; the video says it was available through the website or API, with no weight-release information given.
- Zonos 3 Mini supports zero-shot voice cloning and controls for accent, age, emotion, and delivery. The presenter describes it as about 1.8 GB and compatible with Nvidia GPUs, Apple silicon, or CPU.
- Sofon Real Time is described as full duplex, able to react to tone and laughter, and available in variants with different speed, expressiveness, and voice-cloning support.
Images, robotics, and music27:22
- MiniMax H3 Image is described as generating and editing high-resolution images, with reference-image support; the presenter points to a separate review for details.
- Lingbot Touch interprets tactile sensor data for robot manipulation. Flash Aloha learns tasks from physical demonstrations; the video says 50 demonstrations per task achieved success rates above 80% in the reported tests.
- Ace Music V1 generates songs from prompts or lyrics and supports performance instructions. The presenter notes that Ace Studio can separate vocals and instruments or convert them to MIDI.
Efficient models and decision tools34:14
- Meg Diffusion assigns more compute to important video regions; code was described as forthcoming. The presenter says the approach can also be applied to image generation.
- Muse Ember is presented as a lightweight multimodal model for tasks such as document inspection, code review, and image classification; access was described through its API or Muse Code.
- Ernie 5.5 Flash Preview was available through an API, with weights planned for release. Gemini Decision 1 answers yes/no or multiple-choice questions with confidence scores and was described as available through the Gemini API, including a free testing tier.
Scientific and visual world models37:03
- Nova Atom Light predicts structures and interactions involving proteins, DNA, RNA, and other molecules; the video describes API and platform access.
- Alpha Genome 2 predicts gene activity and the effects of DNA mutations for specified cell types. The video says API access was available for non-commercial research, with commercial access in private preview.
- Iris Z generates images with depth maps and can estimate depth from an image; the video reports model weights and inference code were released. The large model is about 64 GB and the small one about 18 GB.
- Pan 2 predicts how scenes change after actions and can generate clips up to 20 seconds; its described model download is about 86 GB. PiH3 reduces video generation to two or four steps; weights and inference code were described as forthcoming.
- Pix Vid generates video directly in pixel space and supports depth-based editing; code and weights were described as forthcoming.
Tools & resources mentioned
- Google Antigravity — harness used for the Gemini 4 Carbon tests.
- Seedream 5.5 — image generation and editing; sponsor segment.
- Gemini app and Google Search — access points mentioned for Gemini 4 Hydrogen.
- GitHub and Hugging Face — repositories or demos mentioned for several open models, including Budget 3D, Haru 4B, Iris Z, and Pan 2; no URLs are included in the transcript.
- Ace Studio — separate vocals and instruments or convert generated songs to MIDI.
- Fish Audio website/API — access to Live 3.
- Muse Code — agentic harness for Muse Ember.
Notes are AI-generated from the video's transcript and may miss nuance.