Hello!
This week saw a notable rise in the release of open-weights AI models, and as this has become a bit of a political talking-point recently I thought it might be useful to define what “open weights” actually means, for anyone who’s wondering.
There’s a common misconception that models contain copies of all of their training data. In reality, when a model is trained very complex algorithms are used to transform the data into a set of numbers that represent it, in order to calculate answers; these are the weights. When a model has open weights these numbers are given to the public. That means anyone can download the model to their computer (or host it on a server), giving privacy and reduced costs, and enabling custom fine-tuning for dedicated purposes.
Open weights is not the same as open source, where the full source code of a software program is made available for free; the creator of the open-weights model can keep the training data and the source code private. And they can limit commercial use of the model through a licensing agreement, letting you use it for your own purposes, but not resell it or any software that’s based on it (this applies to open source too).
Right, definition over, so let’s get on with it. Hope you’re well!
Made By Google 2026
Google’s annual hardware event saw the launch of four new Pixel phones (11, 11 Pro, 11 Pro XL, and 11 Pro Fold). The phones will feature more Gemini Intelligence, some nice camera software upgrades including Magic Capture which chooses and upgrades stills from a video so you don’t need to switch modes, and new Creator Studio with teleprompter and Storyboard editor.
Also announced: Pixel Watch 5, with on-device Gemini and new health features, Pixel Tag tracker, upgraded Pixel Buds capabilities including dynamic noise cancelling and Gemini audio control, and more connected apps for Gemini (which just passed the 1 billion user milestone).
It wasn’t the most exciting set of announcements, but there were some decent quality-of-life improvements. I wasn’t planning to buy a new phone but with the incentives of a Pixel Superfans discount, a trade-in boost, previous Google Store credit, and six months of Gemini Pro, it turned out that I wouldn’t need to spend a lot to get an upgrade; so I did.
Synthetic Content
LTX-2.5 is an updated video model with native audio that brings improved multishot generation, prompt understanding, and visual quality, including HDR workflow support for filmmaking. ltx.io
It’s not a top-tier model but it’s decent, very fast, and cheap. Here ⬇️ are a few quick tests:
Sand AI’s Magi-2 Preview is a huge (114 billion parameter, for the nerds) unified audio and video generation model with an emphasis on cinematography and character performance. sand.ai
This only appears to be available through a (paid) API so I haven’t had the time to set it up and test it yet. I remember not being hugely impressed by the previous version [TL 377] but it’s been over a year since that was released so hopefully this will be better.
Perhaps the most notable feature of both of LTX-2.5 and Magi-2 Preview is that they’re open-weights releases; with these and the recent releases of MiniMax H3 [TL 440] and FLUX 3 Video [TL 441] this is a golden age for open weights video models.
Seedance 2.5 can now generate native 1080p videos with native 10-bit color, richer textures, and more natural lighting and skin tones. x.com/BytePlusGlobal
It’s rolling out to all the usual creative tools now.
Microsoft released MAI-Image-2.6, with improvements to text rendering, portraits, and commercial design. microsoft.ai
I made this ⬇️ comparison of 2.6 (left) with 2.5 (right); the newer model is more detailed and photorealistic, but perhaps has less… character?
Music
MiniMax Music 3.0 is a music generation model that can produce tracks up to five minutes long, with improved audio detail and song structure. minimax.io
Like the video models above, this has been released with open weights. In my very limited testing (for example this ⬇️) it seems it’s exceptionally good at structure and vocals, but doesn’t really follow my instructions for instrumentation and genre; that might be because I haven’t learned its prompting syntax.
Mureka’s V9.5 music generation model promises more natural and expressive tracks across a range of styles. mureka.ai
There’s no free tier available so I’ve never tried Mureka.Suno introduced Studio 2.0, a major update to its browser-based generative music studio, adding MIDI editing, custom audio plugins, an AI assistant, and more. suno.com
Spotify introduced AI Persona badges, a label which shows that “an artist’s identity may be AI-generated and does not represent a real person“. newsroom.spotify.com
Profile-owners can self-declare, and Spotify will also use automated and human judgement on suspected undeclared profiles. Any profile with a label will be excluded from editorial and algorithmic recommendations by default. I found this interesting:
Listeners have been clear in telling us that they don’t like seeing an artist profile that seems human, only to find out that the persona is AI-generated.
So AI output isn’t judged by whether people like it, but whether they feel ‘cheated’ by it. I wonder how soon, if ever, that will change.
Creative Tools
Multiple updates from Higgsfield: Cinema Studio 4.0, brings video generations up to 30 seconds (powered by Seedance 2.5) plus granular controls for camera motion, lens styles, and character emotions; Marketing Studio adds more presets for UGC ads, posters, product shots, etc; and the ChatGPT Plugin gives access to all of its models plus skills for common tasks including narration and subtitles.
Krea is now available in Slack, giving users access to its creative tools in the messaging app. instagram.com/krea_ai
Getty Images launched an MCP server, making its licensed visual content available to AI agents and workflows. newsroom.gettyimages.com
Even with the general availability of AI images, the requirement for physical photography – whether for events, legal, ethics, or preference — hasn’t gone away. Getty launched its own AI model but it was never any good; licensing and integration seems like a surer bet.
FLORA introduced Fashion Studio, a design platform targeted at apparel brands, letting designers turn sketches into photorealistic garments and virtual photoshoot assets. flora.ai
FLORA added Custom Actions, letting users build bespoke tools such as graphical filters and shaders into their workflow. x.com/floraai
Wan added Agent Skills to its video platform, for building reusable creative workflows. x.com/Alibaba_Wan
Tavus made it easier to deploy its PAL avatars using a widget, an embed, or on its own landing page. x.com/tavus
Google Gemini and Flow now let you toggle off the visible watermark from generated images and video. x.com/GeminiApp
A small but very welcome change.
Assistants
Google released Gemini 3.7 Flash. The new ‘workhorse’ model brings improved performance for more complex (agentic) tasks such as coding and Web design, and in “knowledge-dense fields” e.g. finance and law. blog.google
It’s available now in the Gemini app and Spark, AI Mode in Search, and in Antigravity and AI Studio at a lower introductory price. It’s only three weeks since the last version of Gemini Flash was released; looks like Google is back on it.
xAI released Grok 4.6 “with a particular focus on long-running agents and more ambitious interactive and visual work”. x.ai
It’s available via API, in Grok Build, and Cursor (now officially part of SpaceX).
Grok Bot is an always-on AI assistant for team workflows, with dedicated cloud computers to automate browser and workplace tasks. x.ai
Grok added a Voice connector to generate synthetic voices for memos, podcasts, and more. x.com/grok
Please think twice before using Grok.
ChatGPT added Computer History, turning activity across apps and websites into ‘memories’ and a timeline that ChatGPT and Codex can reference. x.com/OpenAI
The Claude plugin for Chrome can run full Cowork sessions, syncing browser tasks and chat context seamlessly across devices. claude.com
Claude will use invisible watermarking in its text output in future releases. anthropic.com
The link above explains how it works; essentially, certain word choices will form a pattern that machines can detect. It mostly works on large blocks of unmodified text; if you heavily edit the output, or even use a custom writing style, it will be less effective. People are freaking out about this, but Gemini has been doing it for two years and the world hasn’t ended yet.
Meta introduced Muse Glimmer, a smaller AI model that’s designed for agent workflows to run locally on consumer hardware. research.meta.ai
Probably the most interesting thing about this is that, continuing the theme of this week’s newsletter, it’s an open-weights release — and Meta says it will open the weights of Muse Spark 1.2 in the near future. It’s worth reading Mark Zuckerberg’s reasoning behind why this is important: The Future is for Everyone.
For everyone to be part of the future, everyone must have the ability to use superintelligence to improve their lives and shape the world. We will offer free versions that will be accessible to billions of people. For those who want to pay to use more compute, there will be a dynamic auction mechanism that will guarantee that everyone gets the lowest price possible for the intelligence and compute they’re using while also ensuring the capacity is used for whatever people collectively find most valuable.
Alibaba released Qwen3.8-27B, an open weights variant of its latest top-class model (which scores particularly highly on the Arena leaderboard for coding) that’s optimised to run on consumer hardware. x.com/Alibaba_Qwen
I don’t cover the release of every LLM (there are too many!) but this deserves a mention because a) it’s another open weights release, and b) Qwen3.8 is a powerful model being given away for free, more or less; an example of AI being used as a soft power play.
Manus announced its timeline to return to independence following its separation from Meta. Affected users will need to back up and restore recent data, to comply with regulatory requirements. manus.im
Microsoft is removing some of its custom Copilot features including Group Chat, Labs, and the Mico persona. support.microsoft.com
These were all part of Mustafa Suleyman’s efforts at “humanist AI”, which tells you something about the company’s plans for the future.
Social
Meta launched Threads as a native VR app for Quest headsets, letting users browse feeds, compose posts using voice dictation, and navigate with hand tracking or controllers. meta.com
Honestly, I didn’t see this coming. But I suppose that as Meta has adapted Threads for the Meta Ray-Ban Display glasses it wasn’t too much extra development work to do this.
YouTube is changing its Partner Program, raising monetisation eligibility thresholds for new creators and Shorts while expanding Premium Lite revenue sharing globally. blog.youtube