Dear Bedrock, I Love You, But Your Open-Weight Models Are Wearing Last Year’s Jeans

I want to begin by saying that I am a fan. I have a whole gateway built around you. I have dashboards. I have a Terraform module that I named something affectionate and then regretted naming. When people at work ask where the models live, I say “Bedrock” in the voice other people use for “my husband’s cousin who is a doctor.”

So please understand that what comes next is said with love, from a man who opened the Bedrock model list for the 10th time today, hoping for a new open-weight model, and left empty-handed.

Here is the situation. Right now there are open-weight models that are genuinely good. Not good for open-weight. Just good. DeepSeek V4 Flash 0731. GLM 5.3 Flash. Qwen 3.8 27B. They write the code, they follow the instructions, they call the tools without having a small existential crisis halfway through the JSON. They cost a fraction of what the big closed models cost per token, and the people who made them handed the weights to the entire internet like free samples at Costco.

And then I open Bedrock to see which of those I can use without hosting anything, and I find GLM 4.7 Flash. Which is a fine model. I’m sure it went to a good school. But it’s a year old, and in this field a year is roughly the Jurassic period. The newest samples are at a table I’m not allowed to sit at.

Now, I know what you’re going to say, because I have said it to myself in the shower. “Davy, you can host them yourself.” Yes. I can. And I did the math. Self-hosting one of these would cost us about $4,000 (minimum) a month, because we have to book a GPU on demand. And here’s the part that really gets me: if we turn it off, we lose the GPU. Somebody else grabs it. It’s like leaving your seat at a concert to go to the bathroom, except the seat costs four grand and the concert is a language model. So it stays on all night, at 3 am, doing nothing, with the meter running, while I explain to finance why there’s a line item that looks like a small car.

Serverless is the entire point. It’s the beautiful promise that I pay for the tokens I use and think about nothing else. No reserved GPUs, no capacity planning, no cold-start prayers.

And here’s the part that matters more than my feelings. When you run an LLM gateway (and a lot of us do), the whole value is that you can swap models behind one endpoint. Right now the closed models are the only fresh ones in the pantry. That makes the “multi-model strategy” a bit of a bedtime story. If I can’t put a current open-weight model in the same rotation, I’m not choosing between options, I’m choosing between brands of the same decision.

Open weights also do things for a team that no pricing slide captures. You can see exactly what you’re running. You can pin a version and have it stay that version, so nobody rewrites your evals while you sleep. And when the work touches sensitive data, there’s a real calm in being able to point at a model and say “that one, exactly that one.”

So here’s my ask, and it’s a small one. Put the current open-weight models on Bedrock as serverless options, and do it fast enough that they’re still current when they arrive. I know that’s the hard part. These models move like toddlers on a sugar high. But a catalog that’s a year behind makes the newest, cheapest, most interesting option the one thing I can’t reach from here.