Google just made a major open model play with Gemma 4. Cursor is pushing AI coding into full multi- aent mode. Meta appears to be testing a whole wave of hidden models behind the scenes. And TI just showed that a small vision model can outperform much bigger rivals in some of the hardest tasks. There's a lot happening here, so let's talk about it. All right, so Google just dropped Gemma 4, and this isn't just another small iteration. This is a full family of open models, four different sizes starting from 2 billion parameters all the way up to 31 billion. And models are actually derived from Gemini 3. So you're basically getting a slice of Google's proprietary research but in an open format. Now the lineup itself is pretty interesting. You've got the smaller models, the so-called effective 2 billion and 4 billion versions, which are designed specifically for edge devices. So, we're talking smartphones, Raspberry Pi, Jetson, Orurin, Nano, stuff like that. These come with 128,000

context windows, multimodal capabilities, and even audio input on the smaller ones, which is kind of a big deal because it means you can run speech understanding locally without needing cloud inference. Then on the higher end, you've got the 26 billion mixture of experts model and the 31 billion dense model. These go up to 256,000 context and they're targeting workstations and consumer GPUs. The 26 billione is especially interesting because even though it has 26 billion total parameters, only about 3.8 billion are active during inference, which gives it way better latency and efficiency. And that's really the theme here. Google is pushing this idea of intelligence per parameter. So instead of just scaling models bigger and bigger, they're trying to squeeze more capability out of smaller footprints. And according to their benchmarks, the 31 billion model is currently ranked third among open models on the Arena AI leaderboard, while the 26 billione is sitting at number six. That's actually pretty strong, especially when they're claiming

these models can outperform others that are up to 20 times larger. On top of that, the 31 billion model scored 85.7% on the GPQA Diamond benchmark, which is a pretty tough scientific reasoning test, and it ranks third among open models under 40 billion parameters. So, this isn't just marketing. The numbers are actually backing it up. Now, capability-wise, these models handle multi-step reasoning, math tasks, structured outputs like JSON, function calling for agent workflows, and they're fully multimodal, so they can process images, video for things like OCR and chart understanding, and even audio in some cases. They also support over 140 languages and can do offline code generation. And the distribution is wide. You can already access them on HuggingFace, Kaggle, Olama, Google AI Studio, and a bunch of other platforms. They integrate with pretty much everything. Transformers, VLLM, Llama.CP, MLX, NVIDIA NIM, and so on. You can

fine-tune them on Collab or Vert.Ex AI, then scale them into production on Google Cloud. But honestly, the biggest shift here isn't even the models themselves, it's the license. For the first time, Google is releasing Gemma under the Apache 2.0 license. That means full commercial use, full modification rights, no weird restrictions, no termination clauses. You can basically take the model, change it, deploy it anywhere, even on prem, and keep full control over your data and infrastructure. That's a massive pivot from their previous approach, which was way more restrictive, and it's clearly a response to pressure from openweight competitors, especially from China. models from companies like Alibaba and Moonshot have been gaining ground and Google is basically saying, "All right, we're going fully open now to stay competitive." And this is probably going to drive a huge wave of developer adoption. Gemma already had over 400 million downloads and more than 100,000 community variants. And now with Apache 2.0, that number is going to explode.

People can fine-tune these models on their own GPUs, deploy them on devices, build products around them without worrying about licensing issues. You're also seeing strong hardware partnerships here. Google worked with Pixel, Qualcomm, and MediaTek to optimize the smaller models for mobile deployment. So, this is clearly AIMEd at pushing AI directly onto devices, not just keeping everything in the cloud. So yeah, this release is basically Google trying to close the gap between open and closed models while also making sure developers stay in their ecosystem. And that kind of improvement doesn't just help researchers or developers, it also changes what becomes possible in full production workflows. Higsfield just pushed their film making system even further with Cinema Studio 3, which is a major jump from what we've seen before in AI video. They are also sponsoring today's video. Now, Cinema Studio 3 is built like a full AI film studio where you can go from a story idea to a finished scene with visuals, motion, and even audio all handled inside the same system. One of the biggest upgrades is a

physicsaware generation engine, so movement actually follows realworld behavior. Things like collisions, body motion, and action scenes feel grounded instead of that typical AI floaty look. On top of that, there's a cinematic reasoning system where you can feed in reference images and describe what should happen, and the model figures out how to actually shoot the scene. So, instead of manually controlling every frame, you're directing at a higher level. You also get native audio built in, including dialogue, sound effects, and background elements, all synchronized with what's happening on screen, which removes the need for a separate audio pipeline completely. And then there's consistency. Characters, environments, and details stay stable across shots, which is something that used to break constantly in AI video. So, this is really moving from AI generation into something closer to actual production workflows. One important detail though, Cinema

Studio 3 is currently available through their business and team plans. So, it's positioned more as a prolevel system right now. So, start building your own cinematic scenes. Link is in the description. All right, now let's get back to the video. Then there's Cursor. And this update feels a lot bigger than just another version bump. Cursor 3 is really about one thing, making AI coding agents easier to manage in real life. Up until now, a lot of AI coding tools have felt like one assistant, one chat, one task at a time. Cursor 3 is moving away from that. It's built more for the way people are actually starting to work now, where several AI agents can be doing different jobs at once while the developer watches, compares results, and steps in when needed. The biggest upgrades are parallel agents, agent tabs, and a redesigned layout that lets you switch between the coding window and the agent view, or use both together. That sounds simple, though it changes a lot. Instead of getting stuck in one long AI conversation, you can now have separate agents running separate tasks side by side. So, one can be fixing

code, another can be testing something, and another can be trying a different approach entirely. That makes the whole thing feel less like a chatbot inside an editor and more like a proper AI workspace. Cursor 3 also works across a much wider range of setups. It supports local machines, remote SSH sessions, work trees, and cloud environments. So, it's clearly AIMEd at more than just solo developers coding on one laptop. It's trying to fit into bigger, messier, realworld workflows, too. Cursor also moved workree support into a new agents window, which makes juggling multiple code paths feel a lot more organized. And it added commands like /worktree for isolated tasks and slash best of for comparing multiple model outputs, which is useful because more developers now want to see a few different answers before picking the best one. There are some practical upgrades, too. MCP apps now return cleaner and more structured outputs. Large file diffs render faster, and enterprise users get more controls around security and attribution. So,

Cursor 3 isn't just trying to look smarter. It's trying to make AI coding feel more manageable, more scalable, and honestly, more useful once projects start getting bigger. Meta, meanwhile, looks like it has way more happening behind the scenes than most people realized. Testing inside Meta AI revealed several model variants that seem different from the current Llama 4 system powering the assistant. The names spotted were Avocado Mango, Avocado 9B, and something called Avocado TH, which likely stands for think hard. That alone is interesting because it suggests Meta is already much deeper into testing its next models than public updates made it seem. One of those versions, Avocado Mango, apparently showed solid multimodal skills and even made an SVG of a pelican riding a bike, which is a pretty random test, though still a useful sign that it can handle more than plain text. The smaller 9 billion version also seems surprisingly capable for its size. So, if Meta eventually replaces Llama 4 with Avocado in its consumer products, users could notice a

real jump in quality. The awkward part is that Meta still seems to be struggling a bit with timing. Reports say Avocado was supposed to launch in March, then got pushed back to at least May 2026 because internal results were not strong enough against top competitors. There were even claims that Meta discussed temporarily licensing Google's Gemini. And honestly, that tells you a lot about how intense this race has become. These companies are under serious pressure to keep up, and even the biggest players seem willing to consider moves that would have sounded crazy not that long ago. and avocado might not even be the whole story. Another model family called paricado was also spotted in Meta's model selector and this one had three versions too. A regular text model, a reasoning model, and a multimodal version that appears to understand images and video. None of them are public yet. And nobody really knows whether Paricado is replacing Avocado, supporting it, or doing something completely separate. Still, the bigger point is obvious. Meta is testing far more models in the background than most people knew. Two

extra agent modes were also discovered, a document agent and a health agent. That lines up with where the whole AI space is going right now. Instead of one assistant trying to do everything the same way, companies are starting to build more specialized modes for specific tasks. So Meta's platform may eventually become less like one chatbot and more like a collection of focused AI tools inside one product. And finally, TII, the Technology Innovation Institute, dropped something a bit more researchheavy. Though the idea behind it is actually pretty easy to get. They released Falcon Perception, a small vision model with 600 million parameters that is built to understand what's inside an image from normal language instructions. So instead of splitting the job across separate systems, one for looking at the image and another for figuring out what to do, Falcon Perception tries to handle the whole thing in one model from the start. Now, a lot of vision systems still work like toolkits stitched together from separate parts. TI is basically trying to simplify that. Falcon Perception reads

image data and text together right from the first layer, which helps it connect what you ask for with what is actually in the image more directly. The goal here is better grounding, better segmentation, and better understanding of messy visual scenes without needing a huge model. The training behind it was still pretty serious. TI says the model went through about 685 giga tokens of training and was designed to get better at identifying objects, reading visual details, understanding layout, and handling more complex prompts. They also introduced a benchmark called PBench to test harder visual understanding tasks more clearly. On that benchmark, Falcon Perception beat SAM 3 across every category they listed. It did better on simple objects, better on attributes, better on OCR style tasks, much better on spatial understanding, better on relationships between objects, and better on dense scenes too. The biggest jump was in spatial understanding where it scored 53.5 compared to 31.6 for SAM

3, which is a huge gap. TI also used the same idea to build Falcon OCR, a much smaller 300 million parameter model focused on document reading. And for something that compact, the results are pretty impressive. It scored 80.3 on OMOCR, basically matching Gemini 3 Pro at 80.2 and clearly beating GPT 5.2 at 69.8. On Omnidok Bench, it reached 88.64, which was ahead of GPT 5.2 2 and Mistl OCR 3, though still behind Paddle OCRVL 1.5. So even though Falcon OCR is much smaller, it's already strong enough to compete with much bigger systems in document understanding, which could make it really useful for large scale OCR work where speed and efficiency matter a lot. All right, that's it for this one. Let me know what you think in the comments. If you enjoyed it, drop a like and subscribe for more. Thanks for watching and I'll catch you in the next one.