Worlds with memory, new voices and AI models: September 27

A world generator needs to preserve a place when the camera returns. A speech synthesizer needs to change its delivery according to directions in the text. An agent needs to select actions and continue a task through connected applications. The September roundup distributes these capabilities across systems ranging from small specialist models to checkpoints hundreds of gigabytes in size.

The detailed AI Search review of September 27, 2026 covers the demonstrations and announcements examined below. Comparative results are reported with their sources and test conditions; availability refers to the recording date.

Tencent presented WorldCrafter, a controllable world generator starting from text or an image. Its task is to recover previously seen places when the camera returns to an earlier viewpoint. The project page describes implicit 3D-aware memory: the requested view determines which information from past observations reaches the generator. This mechanism requires no explicit depth estimation.

Reconstructed point clouds and camera trajectories appear beside the generated videos. They validate spatial consistency; the review conflates their role with the memory mechanism. Demonstrations include different artistic styles, dynamic scenes and first- or third-person views. The presenter reports base and faster distilled versions and estimates the combined components at approximately 147 GB. That is a file size, not a verified VRAM requirement.

Ming Image 0.1 is intended for visual design, including interfaces, infographics, posters and other images containing numerous text elements and objects. The presenter describes it as an open text-to-image model with six billion parameters. It supports resolutions up to 2K and alpha-channel output, allowing individual objects to be generated against transparent backgrounds and subsequently incorporated into other designs. The main version is called Ming Image 0.1 Design.

A second version, Ming Image Design Layer, accepts a completed flat image. The user specifies which elements to separate, and the model produces multiple transparent layers that can be edited independently. In the Artificial Analysis table shown, the main model ranks first among open-weight models for UI/UX design. The presenter considers it a possible leading open option for designs containing substantial text and many elements. He identifies its license as MIT. The main model files occupy 12 GB and the layer-separation model files 24 GB. These sizes do not establish VRAM requirements.

Mira Scene reconstructs a three-dimensional scene containing separate objects from an image. The system first identifies objects, estimates depth and generates each object's geometry. It then reconstructs positions and sizes to align the scene with the original image. The resulting objects can be edited, imported into Blender and animated. Another illustrated application is using the scene in a physics simulator to train robots. According to the presenter, the code, instructions and dataset have been released. The combined file size is approximately 6.1 GB; VRAM requirements have not been separately verified.

The Flux team, previously associated with image and video generation, introduced FLUX.3 Action for controlling robots. The presenter describes a world-action model with seven billion parameters. Its inputs are a goal, camera images and joint positions. It predicts the next action and what the world will look like after that action. In an example involving moving an object, a robotic arm receives a sequence of motor commands, executes them, observes the scene again and corrects its actions until the goal is achieved. The model is also demonstrated in video games. In the comparison presented with Cosmos 3 Nano, FLUX.3 Action has less than half as many parameters, a higher success rate and greater speed. Weights, documentation and instructions for fine-tuning on a user's robot, game or other device are available; the base model occupies 14 GB.

According to the recording, Google released Gemini 3.8 Flash TTS and Flash Lite TTS for speech synthesis. A voice can be created through a text description. One example is an energetic DJ from Melbourne addressing listeners before introducing music; a short line about a Japanese dragon follows. Another feature is adding metatags within the script to control expressiveness and how particular passages are delivered. The presenter demonstrates their use in conversational and dramatic speech.

In the first extended example, the voice reports that an order was not delivered, asks the listener to wait, updates its status and reschedules delivery. In the second, several lines portray people noticing something large near the edge of their flashlight beam in a forest. They debate whether it is a deer, realize the creature is watching them, initially try to leave quickly and finally decide to run after hearing branches snap. These audio examples illustrate control over delivery and emotion through metatags; they do not claim that real order processing or any action beyond speech generation took place.

Flash Lite TTS is intended for large volumes of speech production at a lower cost. Support for more than 100 languages and dialects is reported. In the presenter's reading of the tests shown, the new Gemini models outperform ElevenLabs on average in synthesis quality and are strong at creating voices from descriptions. Flash ranks second in the Voice Arena blind-comparison table and first in pronunciation robustness. The presenter calls the service inexpensive. He reports access through an API, a programmatic interface for applications to send speech-generation requests, free experimentation in AI Studio and access in Gemini Notebook, which he says was previously called NotebookLM.

Speech synthesis is one stage of localization: our video translation and dubbing project first transcribes the recording, separates speakers and translates their lines, then checks the voice and timing.

Xiaomi completed training MiMo 2.6. The presenter recalls that the previous week the company had shown reinforcement learning in real time, with training runs and progress visible on a dashboard. In the newly presented tests, Pro approaches GPT-6 Astra and exceeds Fable 5. The review also highlights a strong result in DeepSWE. In agentic and knowledge-work evaluations, he also sees performance close to frontier systems, highlighting a CyberGym result above 95%. Demonstrations show the model creating three-dimensional environments and games from requests.

In the Artificial Analysis index shown, MiMo 2.6 Pro becomes the highest-ranked open model, one point ahead of GLM 5.3. The reported cost is approximately 13 cents per task, below GLM 5.3 Flash and DeepSeek 4.1 Flash. The presenter uses these comparisons to describe it as particularly cost-effective. Weights are available through Hugging Face. Pro has approximately one trillion parameters and occupies around 573 GB; Flash has 311 billion parameters and occupies 178 GB. MiMo Desktop, a desktop application for working with the models, was also introduced.

StepFun introduced Step 5 Preview, a multimodal agent model using a mixture-of-experts architecture. Of its 600 billion parameters, 27 billion are active during use; it supports a million-token context. From the programming, agentic and knowledge-work benchmarks shown, the presenter considers it comparable to GLM 5.3 and Kimi K3, sometimes exceeding them. In the Artificial Analysis table it trails GLM 5.3 by one point but costs almost three times less; he also describes closed models as more expensive. One example converts a photograph into a three-dimensional scene. Good financial-research results are mentioned separately. According to the page shown, the model is already accessible through company products and its API, with release of the model itself planned for October 15 at the time of recording.

Another world model is GAE, Geometry-Native Autoencoder. Alongside generated video, it produces a depth map, a camera trajectory through three-dimensional space and a reconstruction of the scene as the camera moves. His explanation is that retaining this representation helps preserve consistency. In one example, the video is displayed beside the emerging trajectory and geometry. Various characters, environments and first-person and third-person views are shown. Installation instructions and training and evaluation scripts have been released; the weights occupy less than 12 GB. VRAM requirements have not been separately verified.

The presenter says Meta previously released a personal agent called Muse. Users can converse with it and request email summaries, shopping, drafted replies and calendar actions. He considers it an easier interface to capabilities resembling OpenClaw but personally feels wary of giving Meta his information. He presents the open-source OpenMuse as an alternative: the agent has a browser, terminal and files and can work on a task for an extended period. Connections to Gmail, Google Calendar and other applications are supported. This is an agent framework that can use a chosen model, either a local model given sufficient computing resources or a paid service through an API key. The presenter identifies an MIT license and getting-started instructions.

Contrastive Language Models, or CLM, quickly score possible actions against the current state. The CLM-v0.1-8B card describes two trainable heads over a frozen Qwen3-8B: one encodes states and the other actions. The model compares their numerical representations and ranks supplied candidates. It does not generate free-form answers, and its probabilities are relative to that candidate set.

The review shows a Dino Run comparison with Jev. The developers report up to nine times lower latency with comparable performance on selected tasks. Separate high DeepSWE and Terminal-Bench verifier scores require fine-tuned heads; they cannot be assigned to the base checkpoint without that step. Weights, the algorithm and launch instructions are published. Fast action ranking and extended solution generation remain different tasks.

The Grok 4.7 segment begins with xAI's claims of lower costs relative to other frontier models. In its own GDPVal evaluation of professional work, the company shows a score exceeding GPT-6 Astra and a few points below Fable 5.1. Evaluations of multi-hour office work and electrical engineering are also presented. The presenter explicitly distinguishes these self-reported results before turning to Artificial Analysis. In that intelligence index, Grok trails GPT-6 and the latest Claude models and ties MiMo 2.6 Pro. Its cost per task in this evaluation is even higher than GPT-6's. The presenter consequently questions the claimed cost-effectiveness and considers its position on the intelligence-versus-cost chart weak. He reports availability in Cursor, Grok Build, through the API and through third-party providers.

The presenter divides GPT-6 into three categories: Astra with the highest scores and price, intermediate Sol and the cheapest, lightest Luna. Sol Max trails Astra and the leading Claude models by a few points in the chart shown but costs over three times less than Astra. Luna sits outside the top five, below DeepSeek and Gemini 3.8, but costs approximately seven cents per task. Comparison with the most expensive Claude gives a difference exceeding one hundred times. The presenter associates Luna with quick, simple tasks that do not require extensive reasoning. According to the announcement shown, Sol and Luna are available in GPT-Work and Codex on all paid plans; free users can access the smaller Luna in the desktop application. They are not yet available in chat at the time of the account.

For Claude Opus 5.5, the presenter offers his impressions after several days of testing, leaving the details to a separate review. He places it among frontier models and praises its work on 3D, design, coding, websites and other interfaces. However, he did not see a substantial advantage over GPT-6 Astra: he considers their intelligence close and Claude's cost per task significantly higher. His personal preference remains GPT-6. Calling Claude models overpriced is the presenter's opinion here.

In a separate announcement, Anthropic claimed progress toward a new enzyme system for gene editing. According to the account, researchers asked Claude to examine a large collection of DNA sequences for unusual reverse transcriptases, enzymes that copy RNA into DNA. Approximately 950 agents used 210 million tokens; after 21 hours, a repeated sequence was noticed that could be connected with the CRISPR mechanism. The function of the finding remains unknown and requires investigation. The presenter emphasizes autonomous searching of a large dataset and narrowing the candidates worth experimental testing. He also relays criticism from biotechnology researchers: CRISPR-like repeats and unusual combinations are common, and most do not prove useful. He therefore allows that the announcement's significance may be overstated; the account does not establish a functioning gene-editing system.

PrimeBot demonstrated T1, which the company, as reported by the presenter, calls a transformable personal robot. It moves between quadrupedal and bipedal humanoid configurations. For carrying objects, the suggested approach is placing a load on its back in quadrupedal mode, then changing configuration for actions requiring a humanoid form. The presenter considers the design interesting and inexpensive, at approximately 20,000 yuan or $2,800. At the time of recording, the company planned to begin shipments in October.

The review reports a company announcement of the first shipment of realistic U1 humanoid robots. The presenter had featured them before and lists their appearance features: faces, hair, clothing and poses intended to look realistic. The news in this segment is the reported start of deliveries for at least the first batch. Autonomous capabilities and the scale of shipments are not specified.

Unitree's Dex-5S robotic hand is human-hand-sized and has 22 degrees of freedom, with separately moving joints allowing flexible finger positions. In the demonstration it holds and manipulates objects, presses piano keys, performs gestures and cuts paper with scissors. Impact torque protection is reported for every joint.

The next demonstration concerns Skild AI and its S1 model for autonomous humanoid soccer. The presenter compares the approach with AlphaGo's self-play training: the system repeatedly plays against itself in simulation. S1's goal is to score. The reported training in NVIDIA Isaac Sim covers the equivalent of more than 140 years of soccer. The presenter explains the choice through the combined demands on balance, speed, ball control, reactions and understanding the game. Initially the robot falls over and behaves clumsily; after reinforcement learning, it can dribble around defenders, shield the ball and shoot. Opponents are previous model versions, so matches become harder as the system improves. Following transfer into a physical robot, some autonomous play against a human is demonstrated. Navigation around people, moving objects and autonomous driving are mentioned as possible further uses of self-play, rather than demonstrated outcomes.

The selection of smaller models begins with Dots3Note from RedNote. The mixture-of-experts model has 208 billion parameters, of which 16 billion are active. From the agentic coding, problem-solving and multimodal-understanding benchmarks shown, he considers performance strong relative to size. His comparison is Kimi K3 with 2.8 trillion parameters, approximately an order of magnitude larger.

The presenter particularly highlights ARC AGI 2, where Dots3Note leads the open models in the table shown. He describes the evaluation as an unfamiliar game environment whose rules the system must understand to win. He connects its difficulty with most models' fixed weights and the challenge of applying newly encountered rules to decisions. The released full version occupies 577 GB and the FP8 version approximately 300 GB. He also describes the model as inexpensive.

OVIS VL Embedding is intended for searching across different data formats. It converts text, images, video and documents into numerical representations in a shared embedding space. The Omni version with three billion parameters additionally accepts audio. Vision variants with two and nine billion parameters handle images and video without audio. In the comparisons shown against similarly sized embedding models, the presenter considers its average performance state of the art. Vision 9B occupies 16.8 GB and Omni 3B 11 GB. An Apache 2.0 license is reported. He presents it as a potentially strong open option for representing different input formats in a common space.

Our knowledge-system project shows how retrieval is tied to answers, with document indexing and links to the retrieved passages.

The model Limite 1B Violetto specializes in difficult mathematics. With one billion parameters, it scores approximately 94% in the competitive-mathematics evaluation shown. The presenter also sees it outperforming models tens or hundreds of times larger in other mathematical tests, with lower computing requirements. He explicitly limits this assessment to its specialization: comparable performance should not be expected in general tasks. Code, models and local installation instructions have been released. He estimates the total file size at approximately 2 GB; VRAM requirements have not been separately verified.

Aikido Altar is a security model based on GLM-5.3. Aikido describes reducing stored weights from 1.51 TB to 488 GB through quantization and then to 328 GB by removing some expert components. Quantization stores numbers at lower precision; pruning reduces the components that must be held in memory.

In an internal rediscovery test of 32 known vulnerabilities, Altar found 23 at least once across three runs. The full parent covered 25. This is combined coverage across three attempts, not recall in each run. Altar’s average per-run recall was 60.4%. The test measures targeted rediscovery within a specified tool pipeline; it does not measure discovery of all unknown errors throughout a program.

The weights are available for self-hosting. At 328 GB, the model still requires a server configuration: Aikido describes four H200s. Compression here reduces a specialist model’s stored footprint while retaining much of its measured capability on a particular task set.

In Project Suncatcher, Google is exploring whether large AI computing systems could be placed in space. The presenter relays the company's rationale: low Earth orbit could provide nearly continuous sunlight and potentially up to eight times as much solar power as terrestrial systems. The first stage is testing whether the components can operate there, for which a prototype satellite carrying Trillium TPU processors is being prepared. A complete orbital data center is not claimed at this stage.

One test concerns launch: during the roughly ten-minute trip, the spacecraft can experience up to 10 g and individual hardware components up to 100 g. According to the account, Google tested the processors with vibration in different directions to simulate launch, and the hardware survived. Another challenge is radiation, which can disrupt electronics and flip individual data bits. The TPUs were tested in a proton beam while actively running AI workloads. Preliminary results, according to the company account relayed by the presenter, showed better radiation tolerance than expected.

For cooling in vacuum, Google is experimenting with heat pipes and radiators that carry heat away from processors and radiate it into space. Coordinated processor operation requires rapid exchange of data between satellites, for which high-bandwidth laser links with extremely precise alignment are being considered. Google plans to test the system with two satellites in 2027. Launch survival, years of radiation exposure, cooling and fast data exchange still need to be established. The prospect of subsequently building solar-powered computing infrastructure remains conditional on solving those problems.

The final substantive topic is Supra Image, an image generator with 100 million parameters. The presenter compares it with Z-image's six billion parameters, making the new model approximately sixty times smaller. He considers the demonstrated images low quality and attributes that to the very small model size. The trade-off is the ability to run on a phone or another small device. The model occupies approximately 417 MB; according to the presenter, its page contains instructions for running it.