A detailed retelling of AI Search (@theAIsearch)’s episode “Weight space programming, science agents, 3D generators, new Gemini Omni, new GPT Image: AI NEWS”, published on 8 November 2026. Original video. The capabilities and availability described below reflect the presenter’s account at the time of the episode; comparative assessments belong to him and the developers he cites.
The episode opens with HairDeck, which reconstructs a character’s hairstyle in three dimensions from a single image. Hair is represented as individual volumetric strands or bundles. Alongside static geometry, the system creates articulated joints so that the hair can move during animation: in the examples, it responds to turns of the head. Colour is retained too, including combinations of several colours. In the comparative tests shown in the video, HairDeck outperforms other approaches in geometric accuracy and generation speed. Code, weights and training scripts are available, with support for NVIDIA GPUs and Apple Silicon. The files on Hugging Face total about 2.45 GB. The presenter suggests that this should make it possible to run the system on most consumer GPUs. The Apache 2 licence permits commercial use.
RelayVSR increases video resolution fourfold. In the comparison clips, the output looks sharper and more detailed, with the model recovering or adding fine features. Its main distinction is how it divides the work between a large and a small model. The large generative model improves only selected keyframes; the smaller model then uses those frames as references when processing the frames between them. As the presenter explains, this division speeds up processing and reduces computational cost. He highlights greater detail and accuracy in the comparison with competing systems. The project has been released, with instructions for running it locally. The total file size is listed as 10.2 GB.
Nitro Fusion 2 generates images in one or two diffusion steps. It is based on Alibaba’s Z-Image; by comparison, Z-Image Turbo normally needs about eight steps. The episode shows photorealistic pictures, artistic styles and compositions containing text or many elements. Its training method assesses both the image itself and its match to the text prompt. This is intended to preserve objects, style and typography even in an extremely short generation process. The code and model have been released, with support for ComfyUI and Diffusers. The full BF16 version takes up roughly 12 GB, and more compressed versions are also available. The presenter links the prospect of running it on consumer hardware specifically to these smaller versions.
NVIDIA Strata models the atmosphere of the entire Earth. It takes temperature, wind, humidity and pressure data and predicts how they change over time. The stated spatial resolution is about five kilometres, with a ten-minute time step. Each predicted state can be fed back into the system to extend the forecast several days ahead. The transformer divides the planet into small regions, processes them individually and combines the results into a global picture. Despite this scope, the model has about 332 million parameters, and the files published on Hugging Face occupy roughly 1.6 GB. Training instructions are available too, and the Apache 2 licence permits commercial use. The presenter argues that Strata models weather and atmospheric processes better than conventional methods and proposes it for weather forecasting and climate research.
FlowHMR reconstructs human movement in three dimensions from ordinary video, without special cameras, motion-capture suits or sensors. Demonstrations show a person kicking a football and doing push-ups. The latest version incorporates physical constraints, aiming for plausible movement rather than poses that merely look similar. The output can animate a character or transfer movement to a humanoid robot: the episode mentions the Unitree G1 and a simulation of its actions. Another proposed use is to connect FlowHMR to a video generator and obtain three-dimensional movement from a text description. The presenter reports an advantage in the accuracy and physical-plausibility tests shown. The code and models are open, and a browser demo is available on Hugging Face; the physical constraints apply specifically to the latest version.
Another NVIDIA project, ReLift, tracks the movement of individual points in three-dimensional space using ordinary video. It continues estimating trajectories when an object is hidden behind another object. In the demonstration, the recovered movement can be viewed from different angles by rotating a virtual camera. Existing models first locate points and estimate their depth; an additional model then corrects errors and predicts the movement of points that are out of view. In the presenter’s assessment, ReLift tracks objects better on average than the methods used in the comparison. Code, weights and training scripts have been published, allowing users to train or adapt the system themselves.
Like its predecessor, Google Gemini Omni 1.2 Flash creates videos up to fifteen seconds long. Inputs can include text, images, video and audio. The presenter draws attention to smooth camera work, plausible movement and more consistent detail in the demonstration scenes. One of the main improvements concerns preserving details from reference images: given pictures of a person and several outfits, the model produces a virtual try-on video that retains the person’s appearance and the characteristics of the clothes. It supports ordinary-language edits, such as replacing a background or object and changing the lighting, as well as transferring motion from one video to a different subject. The episode reports a 30% price reduction: a second of 720p video now costs seven cents. AI Studio, the API and Google Flow are listed as access points.
The next announcement is OpenAI’s GPT Images 2.6. The presenter sees it as a modest improvement on the previous model, retaining text-based editing and reference-image support. A new capability is animated GIF generation. OpenAI says generation is 50% faster. The API offers two variants: flare prioritises speed and can produce an image in as little as six seconds, while sunburst prioritises quality and the preservation of detail across successive edits. The episode also mentions marking regions for precise changes and creating images with SVG vector export, for example for logos and scalable graphics. According to the page shown, Images 2.6 is already available in ChatGPT and Codex across all subscription tiers.
The most extensive research segment covers MIT’s work on weight-space programming. The presenter explains that a neural network’s weights change during training and determine the capabilities it acquires. The usual way to improve a model is further training or fine-tuning, with successive adjustments to those parameters. The researchers test a different approach: can a stronger model inspect another network’s weights and alter them directly to correct its behaviour, without a conventional training cycle? GPT-6 Astra fills that role in the episode. The experiments are described first, followed by the limitations that prevent them from being treated as a universal way to increase intelligence.
In the first experiment, Astra received the weights of four small models trained on different games. It was not given the game titles, yet identified all four correctly, including Pong and Pac-Man. In the second, a small model controlled a car that became stuck on an uphill section. Astra was given its weights and a recording of a failed attempt. After 3.5% of the weights were changed, the car completed the course in every trial performed. The presenter interprets this as a correction to the parameters responsible for the troublesome behaviour. These are results from those particular tests; a general ability to repair arbitrary control systems remains unestablished.
The third experiment involved Alibaba’s small Qwen3.5-0.8B model, with 800 million parameters. Astra was asked to improve its mathematical ability. After it analysed the model and changed a small proportion of its weights, the Math500 score rose from 7.9% to 48%; instruction following also improved substantially. However, the original model had a known defect: it repeated its reasoning, became trapped in loops and exhausted the token limit before reaching a final answer. As the presenter explains, the correction may have enabled it to finish problems it could already solve, without acquiring new mathematical knowledge. He cites an independent analysis in which the unchanged model reached about 44% on Math500 under different settings.
Astra was also doing more than examining hundreds of millions of numbers. In the Qwen experiment, it had Python tools, repeatedly ran the model on training questions, studied its behaviour and tested different changes. The result therefore includes experimental debugging, rather than only a direct understanding of knowledge contained in the weights. A similar attempt to improve Google Gemma, which did not suffer from those reasoning loops, produced no improvement. The presenter says others have reproduced the work, but cautions against expansive interpretations. He regards it as a promising way to find errors and improve models, while a general method for arbitrarily increasing the intelligence of any network is still an open question. The experimental logs are available on GitHub.
Lucid Lyra turns a single image into an explorable three-dimensional scene and identifies individual objects within it. Users can select, extract, move, resize, duplicate and delete them. Another proposed use is to create simulation environments in which robots navigate and interact with objects. A video generator first predicts how the source scene would look from different viewpoints. Another model then combines those views into a three-dimensional reconstruction and separates the objects. The presenter notes an advantage in the object-recognition and separation comparisons shown. At the time of the episode, the code is listed as coming soon.
P-H3 Image repurposes the MiniMax H3 video model as an image generator and editor. It creates pictures from text, changes existing images through ordinary instructions and combines up to eight references in one scene. The examples span photorealism, cinematic scenes, portraits, anime and other artistic styles. The architecture was adapted to still images, and the number of steps was reduced from fifty to eight; the episode reports an approximately fivefold speed-up. It supports resolutions up to sixteen megapixels and automatic expansion of short prompts. Weights are available on Hugging Face in BF16, 8-bit and 4-bit versions, with Diffusers and ComfyUI supported for inference.
OpenRoboKit brings together four open libraries for robot control, simulation and training. RoboKit handles motion computation and action planning: robots in the examples turn handles and cubes, open notebooks, cross obstacles and jump. RetargetingKit transfers human movement from video, motion-capture systems or VR to robotic hands and full-size humanoids. GraspKit helps robots grasp and manipulate objects, train control models and apply them in simulation or on physical hardware. RoboSkills supplies workflows for AI agents, including generating movement, reconstructing simulations from video and improving control autonomously through experiments. The code for all four toolkits and their installation instructions have been released.
MOSS-Live is a real-time multimodal voice system from the team also working on Moss Image. It can receive images from a screen or camera. In the dialogue shown, a user asks it to plan a walk: the system identifies the location as Fuzhou Road, suggests walking from Sichuan Middle Road to the Bund, continuing towards Nanjing Road, having dinner nearby and allowing time to reach the airport. It promises to check journey times. The presenter also describes simultaneous listening and speaking, distinguishing a new question from a brief conversational response, separating the user’s voice from other people and background noise, participating in meetings and searching for information. In another example, a traveller shows the system a location through a camera and asks questions about it.
MOSS-Live assigns conversation and background tasks to separate components. One conducts the real-time dialogue; the other handles complex reasoning and tool calls. This allows the conversation to continue while another task is under way. The presenter gives a positive assessment of the tests shown for spoken interaction, interruptions and tool use. An online demonstration and API are available at the time of the episode. No release of the weights is announced.
Pyramid JiT is an open image generator with about two billion parameters. According to the figures presented, it was trained from scratch with a compute budget of $21,000. Examples include portraits, landscapes and stylised illustrations; the presenter finds the quality unexpectedly high for its scale and budget. The model works directly with pixels, without a separate compressed latent representation, while architectural choices reduce training computation. Its developers call it experimental and acknowledge difficulties with complex scenes, multiple people and precise text rendering. Code and weights are open, and the model takes up 8.74 GB. The presenter suggests that it should fit on most consumer GPUs. The Apache 2 licence permits commercial use.
Cola 3D builds an entire three-dimensional scene from one image, with a separate polygon mesh for each object. It has three modes: automatic detection of every object, selecting objects with points, or supplying segmentation masks. The system processes the scene as a whole instead of generating independent objects and assembling them afterwards. As the presenter explains, this helps retain relative positions, dimensions and spatial relationships. Inference code, training code and a dataset have been published.
Mirage is a small, open computer-vision model. It detects objects and traces their boundaries, reads text and recognises spatial relationships, such as which objects are connected and what is on a table. Users can also supply a set of possible answers for the model to choose from based on the image; the presenter connects this to classification and object counting. A shared component processes the image, while specialised components extract information about objects, text and other details. The model has about 51 million parameters and, after two-bit quantisation, occupies roughly 22 MB. The episode claims that it processes an image on a smartphone in about a tenth of a second; the particular device is not named. Local inference is also intended for laptops, wearables and other consumer devices, and the code and model have been released.
DreamX transfers the movement of a character in a source video to another character. In the demonstration, a human hand manipulating objects is transformed into a robot performing analogous actions. The system can also convert the movement into commands that a physical robot executes, extending the result beyond a new video. A video diffusion model extracts movement information, separates it from the source subject’s appearance and applies it to a different subject. The presenter reports an advantage over some of the compared motion-understanding and transfer methods. The code is promised but has not yet been released.
Zero Real Agent Policy, or ZAP, takes camera video and a text instruction to control a robot directly. According to the episode, its control training took place entirely in simulation, without real-world training data. It starts with a vision-language model that already understands objects and instructions, then uses reinforcement learning on forty relatively simple virtual tasks. When transferred to a physical robot, it also performed new tasks. In the examples, an arm moves a test tube, places a chess piece according to a diagram, holds a pen and draws a cross on a noughts-and-crosses board. The model has 27 billion parameters; in the real-world tests presented, it matched GPT-6 Astra’s success rate while running faster and at lower cost. Only a technical paper has been published so far, with no release of the weights promised.
Tencent Hy4 has gained visual understanding and software-interface capabilities. The presenter shows satellite-image analysis, reconstruction of a house in Blender, creation of a detailed three-dimensional environment in Three.js, financial modelling and stock research. He also mentions reconstructing a three-dimensional mouse embryo from seven million cells in an hour and a half. Based on the agentic coding, three-dimensional modelling and knowledge-work tests presented, he places Hy4 among the strongest open models. He immediately qualifies that assessment: the weights have not been published and are promised in the coming days. For now, access is through Tencent Cloud and WorkBuddy, a free agent environment he compares with Codex and Claude Code.
Hydrogen 1.0 targets scientific work in chemistry, materials science and biology. It has 753 billion parameters and is based on GLM 5.3, with additional training on scientific texts, laboratory records and real research experiments. In one example, the model writes and runs a molecular simulation of polyethylene in a plastic bag, finds an error in its own code, fixes it and continues the calculation. In another, it analyses an NMR spectrum to identify possible structures of reaction products. It also prepares an experimental protocol for a robotic laboratory, checks the instructions in a virtual environment and corrects an error before execution. The training data includes complete research trajectories, including failed experiments.
According to the presenter, Hydrogen outperforms GPT-6 Astra in some of the scientific-reasoning and experiment-planning tests shown. Three versions have been published on Hugging Face: the full model, FP4 for NVIDIA and MX for AMD. Even FP4 takes up about 464 GB, so local inference will require several GPUs, as he notes. The episode’s high assessment of the model’s scientific capabilities concerns those particular examples and benchmarks.
OpenAI’s GPT-6 Cyber specialises in cybersecurity: finding vulnerabilities autonomously and helping fix software. According to the account in the video, it helped discover 81 previously unknown high-severity vulnerabilities in Firefox after earlier AI models had checked it. It also helped develop working exploits for more than nine hundred Linux vulnerabilities. In the specialist tests shown, it substantially outperforms the previous cyber model and, on most of those tasks, GPT-6 Astra. A general public launch is not announced: access is provided to approved security researchers through the Daybreak programme.
Meta is adding real-time voice conversation and expressive avatars to its Muse personal assistant. Suggested uses include holiday planning, discussing ideas and practising for an interview. The assistant can carry out tasks while talking, so the user can keep the conversation going without waiting for a task to finish. The avatar accompanies speech with facial expressions and gestures appropriate to the conversation. According to the page shown, the update is beginning to roll out in the United States in the mobile, web and Mac apps.
Mobius combines generation, understanding, completion and editing of human movement. Given a text description, it creates an animation of a three-dimensional character; given an existing movement sequence, it produces a verbal description. If only part of an action is supplied, the model fills in the missing parts, optionally following a text instruction. Existing animation can also be changed through ordinary instructions. A common architecture handles language and movement, allowing one system to perform both kinds of task. The presenter highlights strong comparative results across several metrics. The code is undergoing internal review and has not yet been made public.
The previously promised weights for Kimi K3.1 have finally been released. The presenter calls it the strongest open model in the agentic coding, general-knowledge and scientific-work tests presented, noting that it approaches some leading closed models. The full version is available on Hugging Face. It is a mixture of experts with 2.8 trillion parameters, of which 104 billion are active at once; the files occupy about 1.56 TB. Instructions cover deployment on different platforms. He stresses that these requirements are closer to enterprise hardware than a home computer, and praises the open release of a model of this size. His suggestion that training probably cost many millions of dollars is speculation rather than a cited financial disclosure.
Baidu Ernie 5.5 Flash has 280 billion total parameters and 14 billion active parameters in its mixture-of-experts architecture. Its context window reaches one million tokens. The presenter highlights strong reasoning, coding and general-knowledge results against open models of a similar size. Hugging Face hosts an FP8 version of about 288 GB and a full BF16 version of roughly 573 GB, along with local-deployment instructions. An important qualification is that this is a base model intended primarily for further training. It should be approached as such, rather than as a ready-made general-purpose assistant.
Ant Group’s Ling 3.2 Flash accepts text, images and video. Demonstrations show it reconstructing three-dimensional objects from images, creating an animated website about London and coordinating agents that turn financial data into an analytical report and presentation. The model has 560 billion total parameters, with 25 billion active at once. In the tests shown, it is competitive with leading open models, including GLM 5.3 Flash and DeepSeek. For now, it is offered through Ant Group’s developer platform. Open weights have not been announced as released: the presenter merely hopes they will follow, because earlier Ling models were opened.
Ripple65B was trained from scratch and has 65 billion parameters. Its developers claim the best agentic coding results among open American models. It uses a dense architecture, with every parameter active during inference, and supports up to one million tokens of context. Ripple was originally built for a platform for designing AI chips: the idea is to improve chips with AI, then use those chips to train stronger models. The team says all research and development cost $32 million; the presenter notes that this is far below the training expenditure of many leading AI laboratories. A 130 GB BF16 version and a roughly 57 GB FP4 version for NVIDIA have been published under Apache 2.
Meituan’s Longcat 2.5 is a multimodal mixture of experts with 1.6 trillion parameters and about 48 billion active parameters. Its stated strengths are coding, image understanding and complex autonomous tasks. Examples include an interactive website with fluid simulation and visual effects, business-data analysis and report creation. Native visual understanding is presented as a major change from the previous version. The presenter notes competitiveness with some OpenAI and Anthropic models, while specifying that these are older generations. An official chat service and API are available; GitHub and Hugging Face releases are still marked as forthcoming.
Thinking Machines has introduced Inkling 2 Small, which despite its name has 594 billion total parameters and 17 billion active parameters. It targets coding, visual understanding and long-running autonomous tasks. In the examples, it creates an interactive city in Three.js with automatically generated buildings, transport and infrastructure, as well as an animation in Blender. In a separate experiment, it autonomously trains and improves another model for time-series forecasting, proposing experiments, evaluating their results and changing the training process. The presenter rates the comparative tests highly and suggests that, once released, it could become the strongest open American model. It is currently available through Tinker; the weights are promised for the following week under Apache 2.
VoxLution generates three-dimensional interiors, including flats, offices and bedrooms, from a category or a simple floor plan. The example rooms are fully furnished, and their geometry includes small details such as table legs and internal cabinet partitions. The system first builds a coarse structure from volumetric elements, or voxels, then progressively raises the resolution and adds detail. The project page still lists the code as coming soon.
The final project, Solver, reconstructs scene geometry and camera positions from several photographs or video frames. Its reconstruction can be refined progressively: new images can be added at any time and in any order. It also supports a real-time video stream, updating the scene as frames arrive. At its core is a recurrent transformer that applies the same computational cycle repeatedly to improve its geometry and camera estimates. The presenter considers the results competitive with current methods and highlights the efficiency of continuous video processing. The repository has been created, but code and models are only promised within the week following the episode.
