Hi tech & AI enthusiasts, this week's highlights:
Alibaba Released the Biggest Free AI Model Yet
MiniMax Gave Away a Video Model That Makes Sound Too
DeepSeek Released a Model 35x Cheaper Than the Leaders
The Lab: How to Run a Cheaper AI Model Inside Codex
AI Image of the Week: The cost of thinking
Let's get into it.
▲ AI MODELS
Alibaba launched Qwen 3.8-Max, its most capable model to date, and is releasing the weights publicly. Anyone can download it and run it without paying per use, though at this size that means renting server time rather than running it on your laptop.
What you need to know
Its biggest model, given away: Until now the top-tier models from every major lab stayed locked behind an API. Alibaba is publishing this one. A smaller version is being released alongside it for people without serious hardware.
It handles text, images and video: Not just a chatbot. You can feed it documents, screenshots and footage and get useful answers back.
Free to run, but not on your laptop: No usage fees once it's on hardware you control, but a model this size needs a serious machine. Most people will rent one from a hosting provider rather than buy it.
The gap is closing fast: Capable models keep getting smaller as well as cheaper. What needs a rented server today has tended to fit on ordinary hardware within a year.
Key Idea
▲ Why it matters: In the past the best models cost money per use, and the free ones were a step behind. But that rule is breaking. If a free model does 90% of what you were paying for, the maths on running AI across your business changes completely. Suddenly automations you shelved because the per-use cost didn't stack up become worth building.
▲ AI TOOLS
Chinese lab MiniMax published MiniMax-H3, a free-to-download model that generates video with matching stereo audio. Up to 15 seconds, up to 2K quality, and it speaks 11 languages.
What you need to know
Video and sound from one model: Dialogue, effects and background audio come out matched to the footage.
Feed it almost anything: Text, up to 9 images, video clips, or audio. It'll use a first frame, a last frame, or both, and keep your characters consistent.
You can run it yourself: The weights are public. There's a paid API if you'd rather not, but the option to self-host is there.
Key Idea
▲ Why it matters: 2 weeks ago a model like this was something you rented from a company like Runway. Now you can download one. For anyone producing video regularly, that's the difference between a per-clip cost that scales with output and a fixed cost that doesn't. The quality isn't quite at the top of the paid tools yet, but it's close enough to matter for social content and first drafts.
▲ AI MODELS
Image from Cline reseach
DeepSeek launched V4-Flash, the cheap, fast version of its flagship. It lands within touching distance of the most expensive models on the market while costing a tiny fraction of what they do.
What you need to know
Nearly as good as the best: On the standard tests for long, multi-step coding work it finishes about 2 points behind Claude's most expensive model. Close enough that for most jobs you would not notice.
Between 35 and 100 times cheaper: Depending on the task, the same work costs a fraction of what you would pay to run GPT-5.5 or Claude Opus. That is not a discount, it is a different order of magnitude.
It beats DeepSeek's own flagship: The cheap, fast version outperforms their bigger, pricier model on agentic work. Usually it goes the other way round.
Key Idea
▲ Why it matters: This is the number that changes budgets rather than opinions. A model that is 2% worse and 97% cheaper doesn't win on quality, it wins on what becomes affordable. Every automation you costed out and abandoned because the per-job price didn't work is worth pricing again. The tutorial below shows how to point the tools you already use at it.
▲ EVEN MORE NEWS
Other big ideas from the past week.
▲ An AI Solved Ten Unsolved Maths Problems for $2,000: OpenAI published ten results on problems mathematicians had been stuck on for decades, found by an unreleased model. The compute cost roughly $2,000. They've been upfront that the model did the thinking, not the humans.
▲ The Model That's Been Quietly Winning on Cost: If you missed it in June, GLM-5.2 from Chinese lab Z.ai is free to download and matches the paid leaders on long coding jobs at around a sixth of the price. It's the one most businesses switch to first.
▲ Moonshot Open-Sourced the Machinery Behind Kimi: Beyond the model itself, Moonshot released the infrastructure it uses to train and serve at scale, giving smaller teams the plumbing that used to be a moat.
The Lab
▲ AI EDUCATION
You don't need to switch tools to cut your AI costs. Codex will happily run a different model underneath, and swapping it takes about five minutes.
What you'll learn:
How to point Codex at DeepSeek using their one-command setup script
How to check it worked and what to do when it doesn't
How to add OpenRouter so you can switch between Qwen, GLM, DeepSeek and others from one account
Which model to reach for depending on the job
🔬 Interested in learning about AI for business? Join The Lab waitlist.
▲ AI IMAGE SHOWCASE
Create this type of image with ChatGPT Images 2
Every story this week was about cost, so this week's image makes that visible.
A clean studio product photograph of two identical server racks standing side by side against a soft neutral grey background. The left rack glows with warm amber internal light, the right with cool blue. A simple paper price tag hangs from the front of each on white string, blank and unmarked. Symmetrical composition, shallow depth of field, soft top lighting with gentle reflections on the floor. Muted palette, photorealistic, shot on an 85mm lens at f/4.Image Usage Suggestion: Pricing page headers, cost comparison blog posts, procurement and tender documents, investor updates explaining unit economics, LinkedIn posts about vendor selection, or any content where you're asking a reader to weigh two options against each other.
💌 Reply to this email with your AI image generation or suggestion of what you'd like to see and I'll feature it in a future newsletter.
▲ SHARE YOUR THOUGHTS
Help shape the content you see here by giving feedback
Have specific feedback or want to get in touch? Reply to this email and I’ll get back to you.
Know someone who’d love this newsletter? Forward it to a friend and have them sign up here.
Thanks for reading, until next time.
Stay curious,
Matt Lok, Editor
Brought to you by Metalabs. Digital marketing consulting specialising in AI. Carousels made with SlideDrift.