Novo robô Raynex G3 com IA, «cérebro» Unitree, ChatGPT Images 2.5, Deepseek V4.1 Flash

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Novo robô Raynex G3 com IA, «cérebro» Unitree, ChatGPT Images 2.5, Deepseek V4.1 Flash · Channel: AI News · 07:33 · Watch on YouTube

Topics: AI agents, AI content generation, AI models, Robotics and autonomous systems

Summary

The video covers Unitree’s H1 1.0 foundation model, which coordinates 64 robot tasks and uses predictive scene modeling, while noting that its capabilities remain demonstrated in controlled settings. It also introduces the Raydex G3 humanoid, highlighting its strength, carrying capacity, and endurance, though no price or release date is announced. OpenAI’s ChatGPT Images 2.5 brings improved editing and new features, while ChatGPT Work adds desktop actions and a data-analysis agent. Finally, DeepSeek’s V4.1 Flash is presented as a more memory-efficient model designed to support long-context tasks.

Key points

Unitree H1 1.0 uses a single 6-billion-parameter model to coordinate 64 tasks, trained on roughly 2,500 hours of real-robot data.
The Unitree system is designed to generalize across different grippers and dexterous hands without retraining a separate model for each.
Unitree’s approach combines embodied reasoning, dynamic region prediction, and action learning; the company has open-sourced code, weights, and datasets.
The video cautions that Unitree’s demonstrations are controlled and curated, unlike unscripted real-world situations.
The Raydex G3 is a 173 cm, 60 kg humanoid with 32 degrees of freedom and stated strength capabilities including a 100 kg deadlift.
OpenAI says ChatGPT Images 2.5 improves detail and editing consistency, adds Sketch and templates, and reduces generation latency by up to 50% compared with Images 2.0.
ChatGPT Work, running on GPT-6 Astra, can operate desktop apps with permission and includes a data agent for analyzing company data in plain language.
DeepSeek V4.1 Flash uses an asymmetric encoder-decoder architecture and is described as requiring less memory and storage while supporting a 1-million-token context window.

Chapters

Start	Chapter	Summary
00:00	Unitree’s humanoid foundation model	The video explains how Unitree H1 1.0 coordinates multiple robot tasks, predicts changes in scenes, and raises questions about how far its controlled demonstrations represent progress toward embodied AGI.
02:40	Raydex G3 humanoid	The Raydex G3 is presented as a strength-focused humanoid, with details on its dimensions, lifting and carrying capacities, battery life, and unannounced pricing and release date.
04:16	ChatGPT Images 2.5	OpenAI’s image model update introduces improved detail and editing, Sketch and templates, and new API models, before the video turns to ChatGPT Work and GPT-6 Astra.
05:53	ChatGPT Work data agent	The video describes permission-based desktop actions in ChatGPT Work and a data agent that can analyze company information and create dashboards.
06:58	DeepSeek V4.1 Flash	DeepSeek’s V4.1 Flash is described as using an asymmetric architecture to reduce memory and storage needs while retaining a 1-million-token context window.

Quotes

And a single instance of it coordinates 64 total tasks — 00:32

real-world AGI doesn’t get to pick its tasks. — 02:40

Raydex G3 stands 173 cm, about 5 ft 8 in tall, and weighs 60 kg, about 132 lb — 03:13

cuts generation latency by up to 50% compared to Images 2.0. — 04:49

while still supporting a full 1 million token window — 07:33

Concepts and entities

Unitree H1 1.0, Unitree, Raynex G3, 9 Ray Group, ChatGPT Images 2.5, GPT-6 Astra, ChatGPT Work, DeepSeek V4.1 Flash, DeepSeek

Full transcript (original language)

00:00 How close is embodied AGI, and would you trust this inside of your home? Today on the AI news, we’re taking a closer look at a brand new humanoid, plus new tech from OpenAI and DeepSeek. But first, these are 10 of the most impressive autonomous executions from Unitree’s newly released >> [music] >> Unitree H1 1.0, which is a new general purpose foundation model that’s built to control humanoid robots across dozens of tasks with a single brain. Rather than training one neural model per skill, which is the way that most in the

00:32 industry still do it. And the model runs on 6 billion parameters and was trained on roughly 2,500 hours of real robot data, not simulation alone. And a single instance of it coordinates 64 total tasks, splitting these between 54 tabletop manipulation tasks and 10 full whole body manipulation tasks. Critically though, it generalizes across hardware too, working with both simple two-finger grippers and multiple different five-finger dexterous hands

01:03 without needing a separate model to retrain for each one. And under the hood, Unitree split the system into two connected pieces. The first is an embodied reasoning model built on top of the Open Qwen 3 VL 4B model and trained on more than 5 million samples covering object detection, spatial trajectory prediction, and multi-image reasoning. And Unitree says this component now leads open-source competitors on seven of 16 spatial and multimodal benchmarks tested and performs comparably to close proprietary models on the rest,

01:36 including recent versions of GPT and Gemini. But something they’re calling dynamic region prediction is where this really gets interesting. So, instead of just reacting to a single still image, the AI model watches optical flow between video frames to figure out which parts of a scene are about to change, then predicts that change ahead of time before it ever commits to a physical action. That future prediction step is then fused with a separate discrete action learning system, which independently encodes gripper position, hand and finger movement, and lower body motion into their own compressed token

02:09 sequences. And all of that gets aligned inside of one shared model, meaning the robot isn’t just imitating a recorded demo, but modeling cause and effect in the scene in front of it before it decides how to act. In fact, Unitree has open-sourced the code model weights as well as training data sets, including a separate 500-hour real-world data set, so that other developers can build directly on top of the same foundation rather than starting over. So, here’s the real question: Is this actual progress towards embodied AGI, or is

02:40 this just a bigger demo reel? One model running 64 tasks with predictive world modeling is a legitimate technical leap, but it’s still tested, controlled, and curated. Whereas, real-world AGI doesn’t get to pick its tasks. So, tell us in the comments what you think a robot walks into an unscripted situation and just handles it with no lab or script. But, while Unitree is racing to build the smartest robot brain, a new company is racing to build the strongest robot body. As 9 Ray Group just unveiled Raydex G3, a new humanoid that’s built

03:13 specifically around raw physical strength. And when it comes to stature, Raydex G3 stands 173 cm, about 5 ft 8 in tall, and weighs 60 kg, about 132 lb, while packing 32 degrees of freedom. Even more notable is that its joints use a custom frameless torque motor design that’s rated for a peak thrust density of 4,500 N per kg. And on raw power, the G3 can perform a 100-kg deadlift, about

03:44 220 lb. It can walk with an 80-kg load on its back, about 176 lb, and it features a dual arm carrying capacity of up to 60 kg or about 132 lbs. And precise dual hand manipulation of up to 30 kg or about 66 lbs are all backed by 44 separate force control units across the body for compliant disturbance resistant movement. And the battery to run it all is rated for 8 hours of continuous use with its joints being tested for over 10,000 hours of

04:16 durability. And there’s no pricing info or release date that’s been announced yet. So, tell us in the comments how much you’d be willing to pay for a Rayna X G3 and what would you like to see this robot do next? >> [music] >> But moving from embodied to generative AI, OpenAI just released ChatGPT Images 2.5 saying this new state-of-the-art image model improves detail, editing accuracy, and speed across a platform already generating more than 3 billion images a week. And OpenAI says the model better preserves people and objects from reference photos, holds up more

04:49 consistently across multiple rounds of edits in one conversation, and cuts generation latency by up to 50% compared to Images 2.0. And two new features ship alongside it. Sketch, which turns a rough drawing into a finished image, and ready-made templates for common formats like posters and product shots. And for developers, OpenAI is releasing two new API models, GPT Image 2.5 Flare, the faster default option, and GPT Image 2.5 Sunburst, built for slower, higher

05:19 precision creative work. In fact, Adobe, Runway, and Midjourney have already integrated the new models into their own products. And Images 2.5 is now live across ChatGPT, ChatGPT Work, and Codex on desktop, mobile, and web. On top of this, OpenAI also confirmed that ChatGPT Work is now running on GPT-6 Astra, positioning it as the company’s strongest model yet for professional work. The company says Astra improves how well ChatGPT Work understands the actual intent behind a request, pulling relevant information from a user’s

05:53 connected apps and files to produce reports, presentations, and analyses that need less manual cleanup and match existing templates and writing styles more closely. The bigger shift, though, is on the action side, >> [music] >> where with permission, ChatGPT Work can now operate a person’s actual desktop apps and browser, clicking, typing, and switching between windows to fill out forms, update spreadsheets and trackers. But, there’s more because OpenAI also launched a dedicated data agent, which allows non-technical employees to analyze company data by asking questions

06:26 in plain language. From there, it can build interactive dashboards inside of ChatGPT or directly inside tools like Tableau, Power BI, or ThoughtSpot, then share findings through Slack or email. And access is permission-based. Administrators choose which data sources are available and to whom, and every query still respects existing table, row, and column level restrictions. And OpenAI says nearly its entire product team and 2/3 of its go-to-market staff already use data agents. And last but not least, DeepSeek has quietly released

06:58 V4 Flash, the smallest model in its newest architecture family, though smallest is relative here. It still totals 552 billion parameters. And like DeepSeek’s other recent models, V4.1 Flash uses a mixture of experts design that only activates a fraction of its parameters per task. Now built on a new asymmetric encoder-decoder architecture that activates 8 billion parameters for processing inputs and 16 billion parameters for generating a response. And for comparison, the previous V4 Flash activated 13 billion parameters of

07:33 its 284 billion total parameters, while the flagship V4 Pro [music] activates 49 billion of its 1.6 trillion. But, the bigger change here is memory efficiency because V4.1 flash needs a quarter of the high-speed memory and an eighth of the storage space that V4 flash required to hold context, while still supporting a full 1 million token window, which should meaningfully cut the cost of running long AI agent tasks. So, like and subscribe and check out this video here for more of the latest in AI and robotics news, and thanks for watching.