New AI models: software work, local DeepSeek, and interactive video

The week, as the presenter describes it, brought several successive language model releases: a DeepSeek model with image support, an Alibaba update, Anthropic’s Fable 5.1, Google’s Gemini 3.8 Flash, Meta’s MuseSpark 1.3 and OpenAI’s GPT-6 Astra. He devotes particular attention to the last of these and considers it the most impressive release in this sequence. Alongside them came projects for interactive worlds, faster video generation, image generation and editing, time series forecasting and weather prediction. The review moves from individual technical developments to large models, before returning to research and visual systems.

The first project, H3-World, turns the MiniMax H3 video generator into something resembling an interactive game engine. In the demonstrations described, the user supplies a text prompt and presses keys, and the system generates video with the corresponding movement or interaction. The developers do not build a separate control system from scratch: key presses become short English descriptions, which pass through MiniMax’s existing language understanding layer. Each command is linked to the specific section of video where the action should happen. The project has already been released, with local running instructions and a training script available.

SolarWM addresses a similar task, converting video models into interactive worlds generated in real time. The presenter highlights its stated ability to maintain consistency for more than an hour, contrasting this with systems whose worlds begin to break down after a few minutes. The framework works with WAN 5B, WAN 14B, LTX 2.5 and MiniMax H3, so the approach is not tied to a single generator. The team assembled 1.3 million video clips, totaling 25 terabytes, and is releasing the dataset free of charge. Models for the listed generators, the data processing pipeline, training recipes and the other parts of the project are also open. The presenter shows pages where the dataset and model checkpoints are available.

Google Research released TimesFM-3, a foundation model for forecasting numerical time series without additional training for a new task. The presenter lists retail, healthcare, weather and stock charts as examples of this type of data. The main upgrade is multivariate forecasting: the model considers several related signals together. With 330 million parameters, it was pretrained on more than a trillion time points. The presenter considers it especially significant that a previously unseen forecasting problem can be given to the model without retraining. According to the Google results he cites, it ranked first among time series foundation models on several benchmarks. The project is open, with instructions on GitHub and weights on Hugging Face. Its stated size is 1.32 gigabytes; the presenter expects a model this small to fit on most consumer devices.

ByteDance’s Lucida creates an editable three-dimensional scene from photographs of a real room, including a messy one. Instead of producing a single fused scan, it identifies separate objects, builds a complete 3D asset for each and returns them to their original positions. Each object therefore remains a separate mesh that can be moved, resized or edited. The process is described in three stages: breaking the room down into objects, reconstructing their complete shapes, including parts hidden in the photographs, and determining their positions in the scene. At the time of the review, only a technical paper had been published.

VideoDeltaNet is another open project for speeding up MiniMax H3. The presenter explains that attention accounts for around 85% of the full model’s runtime. The new version uses a hybrid approach: nearby frames retain the more expensive attention mechanism, which matters for fine detail and motion, while information from distant frames is handled through cheaper linear attention. He likens this to processing nearby surroundings in detail while keeping compressed memories of everything farther away. Combined with optimized computational kernels, parallelism and eight-step distillation, the materials shown indicate a substantial speed increase with little loss of quality.

In comparison videos, the presenter contrasts the original MiniMax H3, another acceleration method called Fast H3 and the new VDN H3. In his visual assessment, Fast H3 noticeably changes the quality, while VDN H3 remains very close to the full model. The cited faster-than-real-time result requires eight NVIDIA B200 GPUs: a 14-second clip is generated in 11 seconds. He explicitly notes that a consumer graphics card will not achieve real-time generation. He compares the speed increase to a Turbo LoRA, but considers the quality substantially better. The model and code have been published, with instructions for local use. Custom ComfyUI nodes have also been created, allowing VDN H3 to be incorporated into an existing MiniMax workflow.

The next open image generator and editor is LLaDA-Image. The presenter shows photorealistic images, as well as posters and infographics containing numerous text and graphic elements. Editing examples change a character’s facial expression and pose, replace parts of an image and its text, and colorize photographs. The model has six billion parameters; the presenter compares its size to ZImage. Two versions have been published: a base version intended mainly for training or higher-quality generation, requiring 50 sampling steps, and a faster Turbo version requiring four. Both have BF16 and FP8 variants. The Turbo FP8 transformer occupies roughly 6.7 gigabytes; the presenter expects it to fit on most consumer GPUs. Download and local running instructions are available.

DeepSeek v4 Flash Vision Experimental adds image capabilities. In the comparison presented, it surpasses the previous v4 Flash without Vision in most cases, and the presenter places it roughly on a par with Opus 4.8. It first became available through an API a week before the release being discussed; it has now been opened for download and free local use through Hugging Face. This is nevertheless a large model, with 305 billion parameters and a total size of approximately 168 gigabytes. The presenter shows a page with running instructions.

Alibaba updated Qwen3.8 Max to Qwen3.8-Max-0902; the designation refers to September 2. The architecture remains unchanged, with 2.4 trillion parameters and a one-million-token context window. The update involves further post-training for coding and collaborative work. The presenter emphasizes that the initial 3.8 Max was released only a few weeks earlier, yet the update substantially improves its results. On some agentic coding benchmarks, the gain exceeds ten percentage points. In some tests of agentic tasks and knowledge work, the new version outperforms Opus 5 and GPT 5.6 Sol. Open weights are not yet available; access is provided through the API on Alibaba’s cloud platform.

Anthropic presents Fable 5.1 as particularly strong at agentic scientific research, coding and knowledge work. In the table shown, it substantially outperforms Fable 5 and GPT 5.6 Sol, while the Fable 5.1 Max version achieves the highest result on an independent leaderboard, above Opus 5 and GPT 5.6. However, the presenter draws attention to its cost: completing a task costs more than 3.7 times as much as with GPT 5.6 Max. His personal experience was unsatisfactory. On a Max subscription, one or two simple prompts exhausted his five-hour limit. When he submitted numerous deep research and medical prompts, he says the service refused to use Fable 5.1 and switched to Opus 5. These limitations, alongside the benchmark figures, underpin his negative assessment of the model’s availability for research.

Google released Gemini 3.8 Flash and 3.8 Flash Cyber. The presenter emphasizes that these belong to the Flash family, yet the results shown for financial and legal tasks, as well as agentic terminal coding, put the standard version ahead of Opus 5 and GPT 5.6 Sol. He also notes strong results in understanding scientific images and charts, long videos, obscure knowledge in Humanity’s Last Exam, bioinformatics and biological research. In a test of long-horizon agentic software development, Gemini ranks first while its average cost is lower than that of other frontier models. There is an important qualification, however: in the same comparison, it produces far more output tokens than its competitors.

Other Gemini rankings give a more mixed picture. In Artificial Analysis, its score is one point below GLM 5.3 and several points below leading Claude and GPT models. The presenter nevertheless cites a speed of up to 348 tokens per second and considers the cost per task reasonable. In Arena, the model ranks eighth, while in LiveBench it does not even make the top ten and falls below Gemini 3.7 Flash. Based on this discrepancy, he voices a suspicion that the model may have been too heavily optimized for benchmarks. Demonstrations include a three-dimensional wizard game, a cross-section and topographic map based on real datasets, and a 3D mechanical keyboard key that can be separated into its components.

Gemini 3.8 Flash Cyber specializes in defensive security. In the vulnerability discovery tests shown, it outperforms GPT 5.6 Sol. According to the developers’ claim relayed by the presenter, Cyber produced 2.6 times as many correct patches for Chrome vulnerabilities as the best commercial models of substantially larger size. Both versions are already available through the API, the Antigravity agentic coding environment, AI Studio and Android Studio. For consumers, the standard 3.8 Flash is available to paid subscribers through the Gemini app, as well as AI Mode in Google Search and Gemini in Google Sheets.

A few hours after Google’s release, Meta introduced MuseSpark 1.3. In the knowledge work and agentic coding tests shown, it performs close to the frontier level and surpasses GPT 5.6 Sol in certain evaluations. The main focus of the release is lengthy, complex assignments with an open-ended goal, numerous files, tools and successive steps. The model is designed to gather information independently, plan, correct its actions and carry the work through to a finished deliverable across different platforms. Improved multitasking is described separately: several assignments within one conversation should be handled without mixing up the instructions belonging to each. At the time of the presentation, MuseSpark 1.3 is available in Muse Code and through Meta’s model API.

The presenter describes GPT-6 Astra as a model for much longer and more complex real-world tasks, including direct control of a computer and cursor. In demonstrations, it turns a circuit schematic into a manufacturable printed circuit board, performs work in Excel, creates a city scene from scratch in Unity, and builds a model of a five-speed car transmission in FreeCAD before animating its moving gears in Blender. Another example involves filling in tax forms and other forms directly in a browser. Later, a house modeled in Blender is shown transformed into a walkable Unreal Engine scene. The presenter uses these examples to illustrate the variety of operations the model can perform through software.

In the tables, GPT-6 Astra achieves leading results in Agent’s Last Exam, OSWorld for operating real computer interfaces, and BenchCAD for reconstructing three-dimensional objects from different views. The presenter associates the largest improvement with computer use. He also shows a test of converting a musical audio excerpt into sheet music, notes design results above Fable 5.1 and Opus 5, and highlights a leading Terminal Bench 4 score on complex terminal-based software engineering tasks. For scientific and medical evaluations, he lists GPQA Diamond, Health Bench and LifeSci Bench as evidence of frontier-level performance. A score of 100% is cited for Exploit Bench; in Exploit Gym, he notes greater performance and efficiency than the previous GPT.

However, independent aggregate rankings do not fully match his impression. Astra ranks second behind Fable 5 in the Artificial Analysis Intelligence Index, and third in LiveBench, also below Fable 5. The presenter says that in his experience GPT-6 performs somewhat better than Fable 5.1, and questions MuseSpark’s fourth-place position. He suspects excessive benchmark optimization and suggests that the intelligence index may need to be reconsidered. In the cost-per-task comparison presented, Astra is substantially cheaper than Claude at a similar intelligence level; this is how he interprets the pricing and performance table.

He discusses ARC AGI 3 in particular detail. An agent enters an unfamiliar game environment and must determine the rules independently, beat the game or advance to the next level. The presenter explains the difficulty for a model in terms of weights that remain fixed after training and the need to apply new information. Most frontier models in the comparison shown score below 10%, whereas Astra exceeds 60% at maximum reasoning effort. With an additional adapter or harness, its result approaches 100%. In a Frontier Math test of extremely difficult unsolved mathematical problems, Astra is the only model scoring above zero; the others score 0%.

Astra also played through Pokémon Fire Red autonomously, with the game available to watch in a Twitch livestream. In the comparison presented, it took less time than previous models. An unexpected detail is that Astra High was faster than Astra Max. The presenter describes access relative to the time of his presentation, a Saturday: GPT-6 Astra should have rolled out to all paid plans, including Plus and Pro. Free-plan users do not yet have access.

Google DeepMind’s WeatherNext 3 updates weather prediction using recent satellite observations. The presenter explains that, rather than relying mainly on weather simulations that may already be several hours old, it receives recent atmospheric data and produces a fresh global forecast every hour. WeatherNext 2 used a 25-kilometer grid and updated every six hours; the new version forecasts variables including surface temperature and moisture at resolutions as fine as five kilometers. The presenter connects the more detailed grid with better representation of conditions in mountains, along coastlines, during storms and in localized rainfall. He cites an improvement of up to 60% over the previous version in a satellite precipitation benchmark.

WeatherNext 3 also forecasts wind speeds at approximately the height of wind turbines and the amount of sunlight reaching solar farms. According to the review, integration into Google Search, the Gemini app, Google Maps, the API and Google Earth Engine begins during the week being discussed. For these products, the presenter separately mentions precipitation forecasts becoming up to 50% more accurate, alongside other improvements.

Another Google Research project is a complete map of a male fruit fly’s brain and central nervous system. The developers call it the largest neural wiring map created. It contains more than 166,000 neurons and approximately 125 million synapses, covering not only the brain but also the ventral nerve cord, which the presenter roughly compares to the human spinal cord. This coverage makes it possible to trace visual, olfactory and auditory signals from different body parts through the brain to the control of movement. To reconstruct it, the nervous system was cut into enormous numbers of very thin slices, photographed using electron microscopes, and rebuilt into three-dimensional neurons and connections using computation and AI.

The presenter emphasizes how laborious this work is and compares the scale: a fruit fly has 166,000 neurons, while a human has 86 billion. He says current technology cannot yet produce an equally detailed map of the entire human brain. Researchers have already mapped a female fruit fly as well, and identify the brain of a larval zebrafish as a next direction. The presenter sees such biological diagrams as a way to better understand the relationship between neural circuits and behavior, and suggests that they might offer ideas for AI systems and robots.

World Labs, founded by Fei-Fei Li, presented a preview of its Atlas world model. It accepts text, images, video and three-dimensional information, combining them to create or reconstruct a 3D world. In one example, an input video and a manually drawn camera path produce up to one minute of coherent 1440p video. The presenter describes the control as pixel-perfect and notes that the scene looks consistent when paused at any frame, including during changes of viewpoint and intense action. In another example, a real place is reconstructed from a handful of photographs. Atlas can also output explicit geometry, such as point clouds or Gaussian splats. The presenter mentions real-estate VR and synthetic data for robots as potential uses. It is currently only a preview; early access can be requested through a form.

InternLumina-U2 combines image generation, editing and understanding. Examples include realistic photographs and oil paintings, changing or removing parts of an image, and accurately following an instruction to rotate only the orange squares. The presenter describes the system as a diffusion large language model with visual understanding. Analysis demonstrations include identifying a neuronal migration defect in a medical scan, recognizing the artist of a painting, and answering questions about a complex ER diagram. These are presented as examples of the model’s capabilities. The code and local running instructions have already been released, but the weights are still awaiting upload to Hugging Face at the time of the presentation.

Viggle Animate replaces a video’s character using a single edited frame. In the process described, an image editor changes the character, for example in the first frame, and the model then extends that replacement throughout the clip while preserving the original movements. The presenter notes that there is no need for separate pose estimation, image segmentation or face tracking. The approach works with characters of different proportions and artistic styles, rather than being limited to realistic humans. It is a fine-tuned and further distilled version of MiniMax H3’s reference-to-video transformer. A full fine-tuned model of approximately 66 gigabytes and a smaller published variant of 2.5 gigabytes are available. Download and running instructions have been released.

The final development is Runway’s GWM Worlds 2. From an initial image or description, it generates an interactive world where users can move in first or third person, specify a character and environment, change the weather, and control the actions of other characters or objects. The output is continuous 720p video at 24 frames per second. The feature the presenter highlights is the absence of a fixed maximum duration. He describes it as an autoregressive diffusion model for video and audio that uses the world it has already created to predict its continuation sequentially. At the time of the review, this is a research preview: it is not available for general public use, and a form must be completed to contact the team.