Which AI Is Best? Start With Where Your Files Go.

Something most people don’t know about me: I’ve always been a gear person. Cameras, computers, whatever’s inside the box, I want to know how it works.

And over the last couple of years, that’s turned into AI. We’ve been building our own software, which we’re launching in 2027, plus some of the hardware and servers that support it. I’ve always been a little hands-on with our technology, but this has meant learning about language models, memory and graphics cards in a lot more detail.

And it matters for the day job too. When we’re planning a project, our clients and our team need straight answers. Which tool fits the job, what it’s going to take, where the files end up. I can’t give them those answers if I don’t understand it myself.

Which means I get asked the same question a lot. Which AI is the best one?

And my answer is always the same. It depends. What are you making? And where are your files allowed to go?

So I want to walk you through it, as simply as I can. This is where things stand in November 2026. Fair warning, half of it will be out of date by spring.

Let’s start with the basics, because there are two kinds of AI model.

The first kind lives in the cloud. When you type into ChatGPT or Gemini, or drop a frame into Runway, your words and images and footage get sent off to their servers, processed, and sent back. These are the most capable models out there. But you can’t download them. You’re renting time on someone else’s machine.

The second kind is open-weight. The maker publishes the model so anyone can download it and run it themselves. “Weights” just means the trained model, one very big file. Run it on your own computer and nothing leaves the room.

That’s the whole private versus cloud thing. Where does the work happen, and who gets to see what you gave it?

For us at Bespoke, the answer changes client by client.

Some clients have policies that say their work can’t go anywhere near a cloud model. Others are totally fine with it. Some will allow one model but not another. And then the models themselves change every few weeks.

At some point, keeping track of all that across our team and every app and tool we use was more than any of us could hold in our heads. So we did the obvious thing and got AI to manage it. Our systems check each client’s policy and flip the switch. Cloud off, and our own models on our own hardware, when the policy says so. Cloud back on when it’s allowed.

It doesn’t replace anyone’s judgement. It just means nobody on the team has to memorise a giant policy spreadsheet on top of doing the work.

Quick detour into what’s inside these things, because it explains a lot of what comes later.

Models are measured in parameters. Think of them as the dials a model tuned while it was learning. More parameters usually means a smarter model. It always means more memory.

Now, most of the big language models today use something called mixture of experts. Picture a building full of specialists. When you ask a question, you don’t wake up the whole building, just the few people who know about that thing. DeepSeek V3 is a good example. It has 671 billion parameters, but only about 37 billion are working on any given word.

That makes it faster. What it doesn’t do is make it smaller. Everyone in the building still needs a desk, so the whole thing has to sit in memory. When you see a model called something like “35B-A3B,” that’s what it means. 35 billion in the building, 3 billion at work.

Image and video models are a different animal. They start with pure noise, like TV static, and clean it up step by step until a picture shows up. Video has to do that for every frame and keep them all consistent with each other. That’s why video eats so much more memory and time than a still.

Okay, the cloud models. I’ll start with images and video, since that’s our world.

For stills, OpenAI’s ChatGPT Images 2.5 is sitting at the top of the image leaderboards right now, and it’s especially good when you need readable text in the picture. Google’s Nano Banana is the one most people reach for day to day. And then there’s Adobe Firefly. It’s not the top scorer, but on eligible plans Adobe will cover you if someone says your output infringes their IP. If you work with brands, you know why that matters.

For video, Google’s Veo 3.1 gets a lot of love for realism, and it makes the sound along with the picture. ByteDance’s Seedance is strong on realism too, and on multi-shot sequences. Runway Gen-4.5 is the one for control. You can paint where you want things to move and direct the camera. And Kling is known for motion and 4K.

You’ll notice one name missing. OpenAI shut Sora down this year, and the reason reported was the cost of the computing power. Think about that for a second. The biggest AI company in the world decided video was too expensive to give away.

Text matters too. It’s still where most of the day-to-day work happens. Briefs, treatments, scripts, research, emails and code. And the big three are neck and neck, swapping places every few weeks.

Anthropic’s Claude Opus 5.5 came out in September, and it’s widely rated the best for coding and for long, careful thinking work. Sonnet 5.5 is the cheaper one most people use every day.

OpenAI’s GPT-6 Astra, also out in September, is the one people point to for the hard stuff. Reasoning, maths, science, and operating a computer for you. Sol and Luna are the cheaper versions, and Luna is what you get on free ChatGPT.

And Google’s Gemini 4 Argon, which went public in October, is top of the big public leaderboard, the one where people vote blind between two answers. In one independent test it also made things up less often than the others. Gemini Flash is the quick, cheap one that can read a huge document in one go.

Now the open models, the ones you can run yourself.

For stills, FLUX.2 [klein] and Z-Image-Turbo both run on a 16GB graphics card, and their licences let you use them commercially. Alibaba’s Qwen-Image is excellent, but be careful, because some versions are non-commercial only. Same with the bigger FLUX.2 [dev].

For video, Alibaba’s Wan 2.2 has a small version that’ll run on an 8GB card. Tencent’s HunyuanVideo 1.5 and Lightricks’ LTX-2 are the other names you’ll hear. LTX-2 is free for commercial use as long as you’re under $10 million in revenue.

But the one I’d want you to know about is MiniMax H3. Back in August it became the first open model to top an independent video ranking. Four to fifteen second clips, stereo sound, and because you have the weights, you can train it on your own footage, characters or style. Sounds perfect, right? Then you read the licence. It excludes the US, the EU, the UK and South Korea.

So that’s the lesson. Open doesn’t always mean you can use it. Read the licence before you build anything on it.

On the text side, most of the strongest open models come out of China. DeepSeek’s V4 is the one everybody points to, and its licence allows commercial use. Alibaba’s Qwen comes in every size from laptop to data centre and it’s great at maths and languages. Moonshot’s Kimi gets praised for coding and fiction, and Z.ai’s GLM for coding agents that work through a task on their own. On the US side, OpenAI’s gpt-oss and Google’s Gemma are smaller and made for normal hardware.

Are they as good as the closed ones? Not quite. On one of the big independent indexes, the top closed models score in the 50s and the best open one scores 46. Close.

Oh, and one thing people get wrong all the time. Using DeepSeek’s own app is not private. Your data goes to their servers, same as any cloud tool. The privacy only comes when you run the weights yourself.

So what does it take to run one of these? This is where the gear person in me gets excited.

Quick rule of thumb for language models. You need about a gigabyte of memory for every billion parameters at 8-bit, half that at 4-bit, plus some room for whatever you’re asking it to read. Images and video need more on top, and the bigger the resolution and the longer the clip, the more they need.

Start with a laptop with 16GB. You can run small language models and the lighter image models. What should you expect? Useful, not impressive. Fine for first drafts and mood boards. Not for anything a client’s going to see.

Step up to a desktop with an RTX 4090 or 5090, that’s 24 or 32GB on the graphics card, and now you’re doing serious image work. The stills can hold their own against the cloud tools on a lot of jobs, especially once you train a model on your own look. Video’s still short and small, around five seconds at 480p to 720p. Great for previz and testing ideas. Not for finals.

Macs are interesting. They share memory between the processor and the graphics, so a Mac with a lot of memory can hold bigger models than most graphics cards. The new M5 Ultra Mac Studio starts at $5,499 with 96GB. Max it out at 512GB and you’re at $22,299, and $11,500 of that is the memory upgrade alone. What do you get for that? The biggest open text models, a step behind the cloud leaders, running at about reading speed for one person. Images work well. Video is where the Mac struggles, because most of the tools get built for Nvidia first.

Next rung up is Nvidia’s pro card, the RTX PRO 6000, with 96GB on a single card. Put four of the slimmer Max-Q versions in one workstation and you’ve got 384GB. Nvidia lists them at $13,250 each now, up from $8,565 at launch, so a full four-card machine lands somewhere around $55,000 to $70,000. That gets you the same big text models, fast enough for a whole team to use at once, plus room for the bigger open video models. This is where local video starts to feel like real work, although most open video still tops out below the 4K you get from the best cloud tools.

And then there’s the data centre, with eight GPUs to a server. Eight Nvidia H200s, about 1,128GB of memory, typically go for around $370,000. Eight B200s, 1,536GB, about $450,000, or roughly $515,000 for Nvidia’s own DGX B200. With that you get the full-size version of every open model out there. And this is the part that surprises people. It still won’t get you Veo or Claude or GPT-6. The top closed models never leave their makers’ servers, no matter what you pay.

So the way I think about it: the Mac gives you the most memory for your money. The Nvidia machines are way faster and can serve a whole team at once. And most studios that need data-centre power probably just rent it by the hour but even that is in high demand these days.

Which brings me to the question I get almost as often. Why is everyone suddenly talking about GPUs?

A GPU, a graphics card, was originally built to draw pixels for video games. It does thousands of tiny calculations at the same time, and it turns out that’s exactly what AI needs. Nvidia makes most of them, and most AI software is written for Nvidia’s chips. That’s how one company ended up at the centre of all of this.

The demand is wild. Nvidia’s data-centre revenue for the quarter ending in January was $62.3 billion, up 75% on the year before. The four biggest cloud companies are expected to spend close to $700 billion on infrastructure this year.

But the real squeeze right now is memory. AI chips use a special stacked memory called HBM, and only three companies in the world make it: SK hynix, Samsung and Micron. HBM takes about three times the factory capacity of normal memory per gigabyte. So every chip that goes to a data centre is memory that doesn’t go into a laptop or a phone. Samsung says its 2026 supply of the newest HBM is already sold out.

And you can see it in the shops. A 32GB kit of regular desktop memory that cost $110 to $140 a year ago was selling for $402 in August. J.P. Morgan thinks memory prices will be up more than 400% from the start of 2024 to the end of this year. And Intel’s CEO has said there’s no relief until 2028.

So yes, running your own models costs more than it did a year ago. That’s a big part of why the cloud is still where most people start.

If you take one thing away from all this, make it these five questions. Ask them about any AI tool, whoever makes it.

Where do my files go?

Are they used to train the next model?

Who at the company can see them?

Could I run it myself if I needed to?

Does the licence let me use the output commercially, where I am?

If nobody can give you a straight answer, well, that tells you something too.

I’ve also put together a longer list of the current models, their sizes, what hardware they need and what each one’s good at. Way too much for here. Just message me and I’ll send it over.

So what are you using for images and video right now? I’d love to hear if anyone reading this is running models in-house yet.