The AI Week That Packed World Models, Open Weights and Data Centers Into Orbit

This week's AI news moved well past the usual model announcement. Within seven days, three world models, four open-weight language models, two robotics headline models, several small on-device models and an initiative to build AI infrastructure in space all arrived together. The picture that emerges is this: remaining production workloads are concentrating in open-weight models, real-time speech and physical world simulation. On the closed-model side, the trend is not a single upward march but a portfolio splitting into distinct cost tiers inside the same family.

Three world models in one week, all chasing 3D memory

The most visible trio of the week consisted of models trying to make video generation backed by three-dimensional memory. WorldCrafter builds an explorable world from a single text prompt or reference image, and while doing so it reconstructs a 3D point cloud , so objects do not vanish when you look away and come back. The arXiv entry for the study, titled WorldCrafter: Consistent Video World Model with Implicit 3D-aware Memory, summarizes the approach. The project page makes the same point in plain terms: exploring a generated world should not mean losing it when you look away. The second tool, MirrorScene, solves an entirely different problem: there is no video here, only a single photograph. It segments the objects, estimates scene depth, generates 3D geometry for each object and then repositions them so the reconstruction matches the original image. That makes every object in a photo individually selectable, and each one can be imported into Blender or wired to a physics simulator for robot training. Unlike many single-image-to-3D methods that return a fused mesh, this approach keeps objects separate, which is what makes the downstream workflow usable today.

The third, GAE, takes its name from Geometry-Native Autoencoder, and the title of the work states the thesis plainly: the learned latent space itself is geometry-native. The work describes a model that does not only generate video but simultaneously produces a depth map, a camera motion trajectory and a 3D reconstruction of the scene. The fact that the total weight stays under 12 GB and that training and evaluation code are released alongside it shows the field is now about reproducible experiments, not just pretty clips.

The open-weight fight is really a unit-cost fight

What reordered the open-weight leaderboard was less the intelligence index than the cost per task . Xiaomi's MiMo-V2.6-Pro took the top spot in the Artificial Analysis (artificialanalysis.ai) open models listing, and the comparison page places it side by side with GLM-5.3 showing a gap of roughly one point. VentureBeat reported the model as the new number one in the open-weight world after DeepSeek. The model sits near a trillion parameters, yet only a small fraction is activated per run. That sparse design is what keeps the capability of a very large model without making it impossible to serve. The Flash version in the same family is smaller and cheaper. Model cards and independent pricing roundups put the Pro tier at tens of cents per task and Flash at around twelve cents. The interesting part is not where the cost was cut but which job has quietly stopped being expensive: automating a two-second utility call no longer requires an average-purpose frontier model. SiliconANGLE (siliconangle.com) covered the launch as an open-sourced model family.

A second Chinese lab, StepFun, also shipped its flagship. Step 5 Preview is a sparse mixture-of-experts model with 600 billion total parameters of which 27 billion are active, a split the Hugging Face (huggingface.co) model card states explicitly. It supports a one-million token context window and prioritizes software engineering and financial research. Artificial Analysis' model comparison page shows Step 5 trailing GLM-5.3 by about one point on the index while running at roughly a third of the cost.

The third open-weight release is the one that caught less attention. dots3-note preview uses a sparse architecture with 280 billion total and 16 billion active parameters and supports a 512 thousand token context window. Its Hugging Face (huggingface.co) model card shows best why a model this size gets close to much larger rivals: the ARC-AGI test , which drops a model into a game environment it has never seen and asks it to learn the rules by itself. Because transformer weights are fixed, such tests are brutally hard for most models; the FP8 release landing at roughly half the size of the full version matters for practical deployment. Single-purpose small models form an interesting undercurrent. Paradigma's Limite 1B is a one-billion parameter model built only for hard mathematics problems, scoring around 94 percent on the competitive math benchmark in its model card, with the whole package under two gigabytes. The week's security-focused release, Aikido Altar, is built on GLM-5.3 and compressed down to under 400 GB at INT4 by removing part of its expert set. The security test result in the AikidoSec model card shows the shrunken version preserves most of the full model's capability. The GitHub repository carrying dots3-note documents the same sparse design and its context length.

Speech, robotics and the hardware of orbit

Google's new text-to-speech models carry the week's audio story. Gemini 3.8 Flash TTS and the cheaper Flash-Lite version can design an entirely new voice from a written prompt, and the model documentation emphasizes studio-grade fidelity, expressive acting and authentic regional accents. Google's blog post confirms both models were announced together and are available through AI Studio and the Gemini API.

Robotics brought two distinct approaches. Black Forest Labs (bfl.ai), known for image and video models, released a robot action model this time: FLUX 3 Action takes a camera frame plus a text instruction and returns the next two seconds of actions as a seven billion parameter open-weight world action model. The Hugging Face (huggingface.co) blog and The Decoder (the-decoder.com) both report it placed first on the RoboLab leaderboard when fine-tuned on the DROID dataset, and the model card puts the weights at around fourteen gigabytes. That size makes it a practical candidate for real-time robot control. The System One idea also found an open release this week. Explaining a class of models that decide quickly and immediately rather than reasoning at length, the speaker points to CLM, released jointly by Stanford and Nvidia. CLM-8B returns a probability score for each option instead of generating a token sequence, and tests show it running up to nine times faster than the earlier JEVA model while matching its success rate in most situations. The reason is simple: many game or vehicle states are just an instant observation plus a choice. The Hugging Face model card also spells out the active and total parameter split. Black Forest Labs highlights the same pivot in its own announcement.

Humanoid robotics was a spectacle-heavy week. UBTECH announced that the first units of its U1 series, designed with realistic faces and hair, have begun shipping, with reporting putting pre-orders above thirteen thousand and prices starting around sixteen thousand five hundred dollars. Unitree introduced a dexterous hand the size of a human hand with 22 degrees of freedom, shown bending fingers, playing piano keys and cutting paper. Skild AI proved something different with its S1 model: a football player that learned by playing against itself in NVIDIA Isaac Sim, accumulating the equivalent of more than 140 years of simulated play. The simulation time reported by GamesBeat is a concrete measure of that scale.

Google's Sun Catcher project is the boldest and least mature entry. Reuters reported that the company plans to launch a prototype satellite carrying Trillium generation AI processors, while Ars Technica (arstechnica.com) noted the satellite will be limited to four chips running in fifteen-minute blocks. Google's research blog describes the aim as equipping solar-powered satellite constellations with processors and free-space optical links. In low Earth orbit sunlight is continuous, which yields substantially more solar energy than ground installations, and four separate obstacles must be solved: launch vibration, radiation, cooling and the laser link itself. The test window reported by Ars Technica shows how early the program still is.

Closed models split into cost tiers

On the closed side, OpenAI introduced the middle and cheap tiers of the GPT-6 family. The announcement describes three tiers ordered from most to least capable, with the mid-tier Sol running substantially cheaper than the most expensive class. The Luna version is positioned at roughly seven cents per task, aimed at fast and simple work. This week the gap in closed-model pricing opened up by two orders of magnitude, and comparison against the cheapest Claude tier shows a difference of several hundred times.

Anthropic's Opus 5.5 was the main closed release of the week. Anthropic's announcement defines the model as the first member of the Claude 5.5 family, and the platform documentation lists a one million token context window with pricing of four dollars per million input tokens and twenty dollars per million output tokens. The speaker's own testing found it extremely strong at 3D design, coding and working with web interfaces, while noting its intelligence is close to GPT-6 Astra and that its per-task cost remains noticeably higher. The most controversial announcement was Claude's discovery in genome data. Anthropic says roughly a thousand agents scanned for repeated sequences over 21 hours and 210 million tokens and presented what they found as a CRISPR-like mechanism. The Verge (theverge.com) report pauses at exactly that point: the finding is not yet verified and researchers in the field say it is not a top priority. Anthropic's own framing emphasizes that the agents searched a massive sequence pool independently and narrowed results to a candidate worth testing, which is the part that actually works. The Verge places that caveat in the body of its report.

Evaluation and the practical takeaway

The week's real headline is not the number of products released but the slide in unit cost. Artificial Analysis' open models comparison shows a twelve cent per-task price is now genuinely reachable, while the same ranking on the closed side still sits in the double-digit dollar range. That turns the question of which model is best into a question of which model suits which job, and it changes how developers actually decide.

A second limitation lies not in the models but in the measurement infrastructure. When a vendor's own benchmarks and an independent evaluation body disagree, a one point difference in the index stops meaning anything. For Grok 4.7 the vendor's own comparisons look favorable while the independent evaluation at artificialanalysis.ai shows the model trailing on the index and expensive per task. A third limitation is what open weights actually mean in practice. Scale now lives in file size : 573 GB for the MiMo family, 577 GB for dots3-note. That table is a reminder that an open-weight model is open to everyone with sufficient hardware, which is a much narrower group than the slogan suggests. Artificial Analysis' open models comparison holds that difference to about one point.

The practical takeaway is simple: stop asking which model is smarter and start asking which model should do which job. For long-running and cost-sensitive work, inexpensive open releases are now good enough. In speech, 3D world building and robot control, the tool chain has been assembled so quickly that the landscape at the end of the year will almost certainly look very different from the one at the start.