What Are the Most Important New AI Models?

TL;DR

Google, Meta, Tencent, xAI, and OpenAI introduced models spanning multimodal reasoning, coding, video generation, image editing, segmentation, 3D reconstruction, and weather forecasting. The most accessible releases include open-source tools such as Depth Anything 3, SAM 3, SAM 3D, and HunyuanVideo 1.5, several of which are presented as practical for consumer-grade GPUs.

Transcript

AI never sleeps and this week has been absolutely insane. We have not one but two new open-source video generators which are state-of-the-art. Google unleashes the best AI model out there, Gemini 3, plus by far the best image editor out there, Nano Banana Pro. But did you know this week XAI also released their top model, Grock 4.1, and then OpenAI ... Read More

Key Insights

• Depth Anything 3 is a 3D reconstruction system that creates spatial maps, depth, geometry, and camera estimates from a few photographs or a video. It remains coherent across multiple rooms and can process difficult footage containing rapid human movement and unstable camera motion.
• Depth Anything 3 is presented as faster, more spatially comprehensive, and more accurate than compared alternatives. Its largest named version has 1.4 billion parameters, while the associated file is described as roughly 7 GB and suitable for many consumer-grade GPUs.
• SAM 3 is an open-source Meta model for detecting, segmenting, and tracking objects in images and videos. Users can select targets by clicking within a frame or entering category prompts, allowing every visible instance of a concept such as zebras to be selected together.
• SAM 3 is reported to outperform leading models across concept segmentation, visual segmentation, object counting, and reasoning benchmarks. On an H200 GPU, it can detect more than 100 objects in a single image within 30 milliseconds, while remaining fast on consumer hardware.
• SAM 3D is a single-image 3D generator that reconstructs selected buildings, objects, and human bodies. It handles unusual postures, predicts full-body meshes, and can reproduce details such as hands and fingers, with separate repositories available for object reconstruction and body reconstruction.
• SAM 3D is shown leading the presented comparisons for 3D shape, texture preference, and human-body mesh reconstruction. The comparison has an important limitation because its Hunyuan 3D results use versions 2 and 2.1 despite newer versions already being available.
• HunyuanVideo 1.5 is an 8.3-billion-parameter open-source generator supporting both text-to-video and image-to-video workflows. Demonstrations show subtle facial expressions, anatomically coherent motion, rendered text, physical interactions, prompt adherence, and controlled camera movements such as panning, pulling back, and rising.
• HunyuanVideo 1.5 is designed for comparatively accessible local use, with the base model requiring at least 14 GB of VRAM when model offloading is enabled. It produces silent five-to-ten-second videos at up to 720p and includes enhancement for upscaling output to 1080p.

Questions & Answers

Q: What is Depth Anything 3 used for?

Depth Anything 3 creates a coherent 3D representation of a physical space from a small collection of photographs or from video recorded while moving through the environment. It estimates scene depth, geometry, and camera positions. The demonstrations include multiple rooms and chaotic action footage, showing that the model can reconstruct spaces even when subjects and the camera move rapidly.

Q: Can Depth Anything 3 run on a consumer GPU?

Depth Anything 3 was released in several sizes for different hardware capabilities. The largest named edition, DA3 Nested Giant Large, has 1.4 billion parameters and is described as offering the most features. Its referenced model file is roughly 7 GB, which the presenter says should fit on many consumer-grade GPUs. The repository provides local installation and execution instructions.

Q: How does SAM 3 select and track objects?

SAM 3 accepts an image or video and lets the user identify objects through visual clicks or text prompts. Clicking can select or deselect a particular region, while a category prompt can select every matching object in the scene. After selection, the model segments and tracks the specified targets, including through crowded, fast-moving, and visually chaotic video sequences.

Q: How fast and accurate is SAM 3?

SAM 3 is presented as outperforming other leading systems across concept segmentation, visual segmentation, object counting, and reasoning benchmarks. For a single image, it can reportedly detect more than 100 objects in 30 milliseconds when running on an H200 GPU. The presenter also characterizes it as very fast on consumer GPUs, although no consumer-hardware timing is supplied.

Q: What can SAM 3D generate from a single image?

SAM 3D can turn a selected region of one image into a three-dimensional reconstruction. Users can choose buildings or individual objects, and multiple items from the same photograph can be reconstructed separately. A specialized body system predicts human meshes from uncommon poses and postures, with demonstrations showing accurate body positioning as well as detailed hands and fingers.

Q: Is SAM 3D better than other 3D model generators?

The presented benchmarks place SAM 3D ahead of compared generators in shape quality, texture preference, human-mesh reconstruction, and user win rates. However, the comparison against Hunyuan 3D uses versions 2 and 2.1, even though versions 2.5 and 3 had already been released. That outdated comparison makes the claimed advantage less conclusive for the newest competing systems.

Q: What are the hardware and output limits of HunyuanVideo 1.5?

HunyuanVideo 1.5 contains 8.3 billion parameters, and its base model can run with at least 14 GB of VRAM when model offloading is used. It creates videos lasting from five to ten seconds and natively outputs resolutions up to 720p. A video super-resolution enhancement can upscale results to 1080p, but the generated clips do not include sound.

Q: What kinds of videos can HunyuanVideo 1.5 create?

HunyuanVideo 1.5 supports both text-to-video and image-to-video generation, with an uploaded image serving as the starting frame for the latter. The examples demonstrate facial emotion, figure skating, extended limbs, neon-sign text, a hand crushing a soda can, and cinematic scenes. It also follows specified camera actions, including downward pans and rising pullback shots.

Summary & Key Takeaways

• Depth Anything 3 reconstructs coherent 3D environments, depth, geometry, and camera positions from collections of images or moving video. It reportedly handles large spaces and chaotic scenes while outperforming competing systems in reconstruction and camera-pose accuracy. Multiple model sizes, local installation instructions, and interactive demonstrations are available through its released repository and project page.

• Meta expanded its Segment Anything family with SAM 3 and SAM 3D. SAM 3 detects, segments, counts, and tracks prompted objects in images and videos, while SAM 3D converts selected image regions into textured 3D objects or human-body meshes. Both releases provide repositories, models, instructions, and options for local experimentation.

• HunyuanVideo 1.5 is an open-source, 8.3-billion-parameter video generator supporting text-to-video and image-to-video creation. It produces silent clips lasting five to ten seconds at up to 720p, with enhancement to 1080p. The broader roundup also covers Kandinsky 5.0, Grok 4.1, Gemini 3, Nano Banana Pro, and GPT-5.1-Codex-Max.