AI never sleeps, and this week has been completely insane. We have real-time interactive worlds generated from a single image, AI systems that reconstruct moving people in 4D, tiny local voice-cloning models, open image generators capable of native 4K output, and humanoid robots that are starting to run, jump and play tennis at genuinely ridiculous levels.
For Canadian businesses, startups and IT leaders, the main story is bigger than any single model. The AI stack is becoming more open, more multimodal and more physically capable. These systems are no longer limited to producing text or pretty images. They are editing video, operating inside business workflows, generating robot training data and learning physical tasks from a human demonstration.
The pace is accelerating. The real question for organizations across the GTA and the wider Canadian tech ecosystem is not whether AI will change operations. It is how quickly companies can identify the useful tools, build responsible processes and turn this wave into an advantage.
Evoke
Evoke is one of the most interesting releases of the week because it moves AI video toward something much more interactive. It is a fully open-source model that can generate an explorable world in close to real time from an input image and joystick movements. Move the joystick, and the generated video responds nearly instantly.
The model appears to understand a remarkable range of scenarios. It can handle snowmobiles, kayaking, scuba diving, rock climbing and other forms of movement through an environment. It can also introduce events with text prompts, such as adding balloons to a scene or producing aurora lights in the sky.
That makes Evoke feel less like a traditional video generator and more like a promptable world model. Its potential business value is huge, particularly for simulation. A hospital, care facility, industrial site or emergency response organization could potentially create synthetic operating scenarios that are difficult, costly or unsafe to collect in the real world.
Evoke is reportedly a 14-billion-parameter model that achieves its speed by generating video in only three steps. Sessions can continue for hours. The final model weighs roughly 57 GB, so a serious GPU is required, but its Apache 2 licence makes the release strategically important for developers who want fewer restrictions.
4DAnyone
4DAnyone takes a single video of a moving person and converts that person into a representation that can be examined from any angle. Think of it as a moving 3D model, rather than a static reconstruction. Technically, it produces a 4D Gaussian Splat that captures both the character and motion over time.
The workflow is clever. The system extracts a 3D skeleton from the source video, uses that skeleton to guide the creation of new camera viewpoints, then reconstructs the character from those perspectives into a coherent animated model. Compared with earlier 4D approaches such as RecamMaster and Trajectory Crafter, the results are described as more detailed and more consistent.
This could matter for digital production, training, simulation and virtual product experiences. A Canadian retailer, sports organization or industrial training provider could eventually capture a short motion sequence and repurpose it into a navigable asset for a digital environment.
The model is around 12 GB, making local use more attainable on mid-range to high-end GPUs than many of the massive video models now arriving.
SenseNova U1.5
SenseNova U1.5 is an open-source image generator and editor that stands out for two reasons: native 4K generation and end-to-end work in pixel space. It can generate photorealistic images, posters and detailed infographics with many elements. It also supports natural-language editing, including changing poster text, selectively altering objects and following instructions drawn directly onto an image.
Most image and video generators operate through latent space, a compressed representation that must later be decoded back into pixels with a component such as a VAE. SenseNova skips that decoding stage. Historically, direct pixel-space generation was too computationally expensive, but this model is designed to make it practical at 4K resolution.
That is a major development for creative teams. High-resolution marketing assets, complex product collateral and presentation-ready visual material are exactly where image AI often hits quality limitations. SenseNova points toward a future where the workflow gets simpler and the output gets closer to production quality.
The trade-off is hardware. At around 50 GB, the model needs a high-end GPU. Its Apache 2 licence, however, means organizations can explore it with relatively minimal licensing restrictions.
Bernini v2
ByteDance has released Bernini v2, an open-source omni-modal video editor. The idea is straightforward and powerful: take an existing video and edit it with natural language. Add characters, remove objects, alter the colour tone, replace backgrounds, shift camera perspective or insert an object based on a reference image.
That kind of editing could radically shorten content production cycles. Imagine turning one video shoot into multiple localized campaigns, changing product staging without reshooting, or removing an unwanted object from a corporate video with a text instruction.
There is one very practical catch. Bernini v2 is enormous, weighing roughly 180 GB for the model alone. Running it also requires additional components such as a VAE and text encoder. That is a significant barrier for most local deployments, and it explains why the previous version has not become a mainstream open-source favourite.
It is impressive research, but companies evaluating AI video editing should separate technical possibility from operational accessibility. Powerful hosted tools may remain the more realistic option for many teams.
Ornith 1.5
Ornith 1.5 is a family of open models built around an ambitious concept: self-improvement through a closed loop. Rather than relying only on humans to create new tasks and training data, the system proposes problems, develops the scaffolding required to solve and verify them, generates solutions and uses the results as reinforcement learning data.
In other words, the system is attempting to create progressively harder challenges for itself. That is especially relevant for agentic coding and knowledge work, where success depends on handling multi-step tasks rather than simply answering isolated questions.
The family includes a 9-billion-parameter dense model, a 35-billion-parameter mixture-of-experts model and a 397-billion-parameter mixture-of-experts model. The largest model performs strongly across benchmarks such as Terminal Bench, SWE Bench, Deep SWE and Frontier Bench. It reportedly surpasses GLM 5.2 across these evaluations despite being smaller, and it approaches the performance of closed-source frontier systems.
Benchmark scores should always be treated carefully, especially when results may favour selected comparisons. Still, the accessibility story is compelling. The smallest 4-bit GGUF option is under 6 GB, making local experimentation feasible even on lower-end hardware.
Audio8 TTS 0.1B
Audio8 TTS is tiny by modern AI standards. At only 0.1 billion parameters, it can take a short voice sample and generate new speech that closely resembles the original voice. It is also multilingual, with demonstrations spanning English, Chinese, French and Spanish.
The exciting part is not merely that it performs voice cloning. It is that the complete package is only around 1.7 GB. That means the model can fit on many consumer devices, opening the door to local deployment rather than forcing every voice interaction through a cloud service.
For Canadian organizations operating in bilingual or multilingual environments, local TTS is particularly interesting. It could support internal prototypes, accessibility experiments, multilingual communications and low-latency voice interfaces. Of course, voice-cloning capability requires clear consent, governance and safeguards. The smaller and easier these models become, the more important it is to establish policies before deployment.
Hubspot Codex prompts
OpenAI Codex is not just another browser-based chatbot. It can work directly with files on a computer, complete tasks and save finished outputs back to the machine. That distinction matters because the real business opportunity with AI is not simply generating answers. It is reducing the repetitive work that eats entire weeks.
HubSpot’s free Codex prompt guide focuses on five practical use cases. One prompt creates a daily work brief by reviewing calendars, emails, messages and open follow-ups, then organizing priorities, meeting preparation and decisions requiring attention. Another creates a manager-ready weekly summary from scattered meetings, documents, product updates and conversations.
The most useful concept may be reusable skills. A process that worked once can be captured as a repeatable workflow, rather than explained from scratch every time. For executive teams and operations leaders, that is the pathway from occasional AI experimentation to repeatable productivity gains.
GeoWeaver
GeoWeaver tackles a difficult computer vision problem: reconstructing a coherent 3D scene from a long video. Reconstructing from a few frames is one thing. Maintaining correct scale, depth and camera position over a long sequence is much harder, because small errors accumulate and the digital world starts to drift apart.
GeoWeaver divides video into manageable chunks, estimates depth and camera position for each section, then progressively stitches those sections into a globally consistent scene. It uses nearby frames, overlapping views and long-range matches between distant portions of video to keep everything aligned.
The reported outcome is better camera trajectories, cleaner point clouds and lower average error than competing approaches. Only a technical paper has been released so far, but the implications are substantial. Construction, infrastructure, real estate, mining and industrial operations all depend on accurate spatial understanding. Reliable video-to-3D reconstruction could eventually make existing video archives far more valuable.
Qwen Video Edit
Qwen Video Edit applies natural-language image editing to video. It builds on Qwen Image Edit and integrates that capability into a video-generation workflow using Alibaba’s Wan technology. The result is frame-by-frame editing of an existing video through text instructions.
This is another sign that the old separation between image tools and video tools is disappearing. If an image editor can preserve enough consistency across frames, it becomes a video editor. That could simplify workflows for creators and marketing teams that need to make targeted modifications without rebuilding a scene from zero.
The code and training scripts are available, but the full setup is around 41 GB and requires a high-end GPU. It is also entering a crowded field where systems such as MiniMax H3 already offer video-editing capability. For technical teams, Qwen Video Edit is a useful open workflow to study. For everyday users, the question remains whether it delivers enough advantage over easier alternatives.
Paxini data glove
Paxini Tech’s PX Cap Pro data-collection glove is built for one goal: capturing human hand data that robots can actually learn from. The lightweight glove combines tactile sensors across the fingertips and palm, a wide-angle wrist camera and precise angular encoders that track finger joints accurately even in magnetically noisy environments.
The tactile sensors can measure force down to 0.1 newtons. That level of sensitivity is critical for dexterous tasks where a robot must understand not only how a hand moves, but how much pressure it applies.
Businesses interested in physical automation should pay attention. Employees could wear these gloves while performing tasks such as tying ribbons, packing boxes, handling balloons or conducting lab work. The resulting data can become a training resource for robots. This is a much more direct route to automation than attempting to describe every physical motion with code.
MX01
ArcShell Robotics has demonstrated a real-life transformer robot called the MXD1. It can operate as a bipedal humanoid, transform into a quadruped, attach to a drone for air transport and is expected to gain a wheeled form as well.
It is undeniably cool. It is also worth being sceptical. Every transformation adds mechanical complexity, integration challenges and more potential points of failure. A robot designed to do everything can sometimes become less effective than purpose-built machines designed to do one job exceptionally well.
Still, the MXD1 highlights a serious design question for robotics: should machines adapt their form to the environment, or should organizations deploy specialized fleets? The answer will depend on the setting. Disaster response and remote inspection may reward versatility. Warehouses and production lines may reward reliability, simplicity and repeatability.
The most important part
The World Robot Conference in China featured an explosion of highly humanlike social robots, including singing androids, lifelike torso models and an elf-themed bionic robot with an articulated body. These demonstrations are striking because facial expressions, blinking, eye movement and head motion are becoming increasingly natural.
The most important part is not the novelty of robot characters. It is the underlying convergence of expressive interfaces, articulated bodies, speech systems and increasingly capable AI control. Social robots are moving beyond static heads and scripted gestures toward machines that can plausibly interact in hospitality, entertainment, education and customer-facing contexts.
There is still a wide gap between looking lifelike and delivering dependable, useful work. But the hardware is improving, and the social expectations around human-machine interaction are changing with it. Canadian organizations exploring public-facing robotics need to think now about privacy, accessibility, brand trust and what kinds of interactions are genuinely useful.
Qiji horse
Dax AI’s Qiji cyber horse is a quadruped robot designed for rough terrain. People can ride it, and the X1 version is priced at approximately US$40,000. A wheel-legged XS version costs around US$53,000 and is intended to carry a person across steep slopes, gravel, mud, snow and ice.
The system has a stated payload capacity of 300 kilograms, a range of 40 kilometres and a top speed of roughly 10 kilometres per hour. A wheeled variant can reportedly reach 40 kilometres per hour.
Humanoid robots get most of the headlines, but legged mobility may deliver earlier value in situations where wheels fail. Canadian applications are easy to imagine in remote terrain, winter conditions and resource-sector operations. The practical question is not whether robot horses will replace vehicles. It is whether hybrid mobility platforms can safely reach places where conventional transportation struggles.
Humanoid games
The World Humanoid Robot Games in Beijing signal a new phase for the sector. Opening rehearsal footage showed large groups of robots, including Tiangong, Booster and Galbot models, marching across a track in what genuinely resembles an Olympics for robots.
These competitions are more than spectacle. They create public benchmarks for mobility, balance, control, perception and autonomy. A robot that can operate under pressure in a competitive setting is still far from being ready for a workplace, but the events make progress visible and comparable.
The speed of improvement from year to year is becoming impossible to ignore. Robotics is no longer advancing only in isolated research labs. It is becoming a high-profile engineering race with real commercial implications.
Superman
Unitree has previewed a humanoid robot called Superman, and the performance claims are wild. It can reportedly achieve a standing high jump of approximately two metres, above the human standing high-jump record of 1.8 metres. It also has a stated top speed approaching 12.7 metres per second, edging past the fastest recorded human sprint speed of 12.4 metres per second.
The machine has reportedly been in development for only a little over three months. If accurate, that makes the rapid progress even more remarkable.
However, raw athletic performance is only one dimension of robotics. The robot’s demonstrations also show it struggling to slow down, sometimes crashing into a barrier or wall because stopping remains difficult at those speeds. That is a perfect reminder that capability and control must advance together.
Sprinting robots
Superman is not the only robot moving at absurd speed. Honor has demonstrated its Lightning robot outpacing a human runner, while Tiangong has also shown impressive sprinting performance. These are early demonstrations, but they reveal a major shift in dynamic locomotion.
Fast running requires much more than powerful motors. A robot needs balance, rapid perception, trajectory planning, foot placement and the ability to recover from disturbances in milliseconds. It also needs to stop safely, which remains a very obvious challenge.
For business leaders, the takeaway is not that humanoid robots will suddenly replace athletes. It is that the mobility barrier is falling quickly. Once robots can move quickly and reliably through human-designed environments, the set of feasible applications in logistics, inspection, emergency response and security expands dramatically.
Tennis robots
Humanoid robots are now playing tennis, and that is far more difficult than it sounds. A system called Adapt transfers movement data from real tennis matches onto a Unitree G1 robot. The machine can rally and serve autonomously.
Galbot has also demonstrated autonomous tennis play during the World Humanoid Robot Games. To do this, the robot must track the ball, move into position, calculate timing, apply the appropriate force and control spin in a split second. Tennis involves topspin, backspin, body movement, anticipation and constant decision-making under changing conditions.
These are not just sports demos. They are compact demonstrations of full-stack robotics: perception, planning, locomotion, manipulation and real-time control. A machine capable of returning a tennis ball is learning the same broad categories of skills required for many dynamic physical tasks.
Deepseek vision
DeepSeek has released DeepSeek V4 Flash Vision Experimental, extending its V4 Flash model with vision capabilities. The model can analyze images, video and documents while retaining the performance profile of the text-only version.
Reported benchmark results show strong agentic coding performance, including results that match the closed-source Opus 4.8 system on several evaluations. Its Deep SWE result also improved by nearly five points in less than a month.
For enterprise teams, multimodality is the key development. Most business information is not pure text. It lives in invoices, product images, scans, charts, reports, dashboards and video. A model that can reason across those formats can become far more useful inside real operational processes. DeepSeek V4 Flash Vision is currently available through an API, so it is immediately relevant for teams building AI-enabled applications.
Happy Shrimp
Happy Shrimp is a new AI music generator that produces remarkably polished output. Users can describe a song’s style, provide lyrics or select an instrumental option. The generated tracks are clean, dynamic and surprisingly coherent, placing it among the most impressive music-generation tools currently available.
This is likely from the same lab behind Happy Horse, and it continues the broader trend: creative AI is becoming capable enough to produce usable first drafts across every media format. Images, video, speech and music are converging into one increasingly accessible production layer.
For Canadian marketing teams, agencies and startups, the opportunity is rapid ideation. Custom background music, concept jingles, internal presentations and experimental campaign materials can be created quickly. The responsibility is ensuring that usage aligns with brand standards, rights considerations and ethical content practices.
Comfy MCP
ComfyUI has open-sourced Comfy MCP, an agentic connector that allows an AI agent to interact directly with a ComfyUI installation. Instead of manually connecting nodes across a complex visual workflow, users can instruct an agent in natural language to generate content using available models and workflows.
What makes Comfy MCP especially useful is its awareness of the local environment. It can understand installed models, custom nodes, GPU hardware and system capabilities. An agent could be asked to generate a video with a particular model, then automatically construct the appropriate workflow without requiring manual node manipulation.
This is a major usability shift. ComfyUI is extremely powerful, but its node-based interface can be intimidating. Agentic control makes advanced local AI workflows more accessible, potentially allowing creative and technical teams to focus on outcomes rather than configuration.
Gen 1.5
Gen 1.5 is a robot foundation model aimed at generalization. Its core breakthrough is one-shot learning: show the robot a task once, and it can sometimes attempt to repeat that task without additional training.
A robot foundation model acts as the machine’s brain. It processes visual input, sensor information and natural-language instructions, then outputs movement trajectories at up to 100 times per second. In Gen 1.5’s one-shot setting, the model can learn from a demonstration lasting only three to 12 seconds.
The results are promising but not perfect. Across 10 diverse tasks, the system achieved a 59% success rate after one demonstration. With roughly five minutes of additional data per task, the success rate rose to 83%.
These are relatively short and simple tasks, and 83% is not enough for many production environments. But the direction is extremely important. The long-term vision is a robot that can be shown a task by an employee, rather than programmed from scratch by a robotics specialist.
Avo
NVIDIA’s Avo framework delivers perhaps the biggest strategic lesson of the week: improving the system around an AI model can unlock enormous performance gains without changing the model itself.
Avo is an agentic harness that enabled Claude Opus 5 to score 100% on ARC-AGI 3, a benchmark where AI systems enter unfamiliar game environments with no instructions and must infer goals and rules through trial and error. Humans find these problems relatively easy, while leading AI models generally perform poorly. Claude Opus 5 on its own reached 30%.
With Avo, it reportedly reached 100% across all 25 environments and completed all 183 levels. The evaluation used the public dataset, so the result may be inflated. Still, the principle is massive: the model is only part of the product. Planning loops, memory, tools, verification and orchestration can materially change what an AI system can accomplish.
That is the urgent message for Canadian business technology leaders. Do not evaluate AI only by comparing model names or parameter counts. Evaluate the complete operating system around the model: data, tools, safeguards, integration, workflow design and human oversight. That is where the next competitive advantage will be built. Is your organization building prompts, or is it building an AI system that can actually do work?
