This is more exciting than k3, IMO. Dsv4 models are extremely cheap to serve. Improving their capabilities has lots of downstream effects, as it becomes "good enough" for more and more tasks.
DS was serving the pro version at extremely low prices for a long time, and they've had integrations with opencode & other providers, so they likely gathered a lot of data from real developers doing real tasks (on openrouter they were labeled as such). Now they can use those live scenarios to further post-train their models and improve them further.
Can't wait to see if distilling k3 into dsv4 brings additional improvements. Anyway, having fast cheap models getting better is great for the community. Especially since these don't "go away" on a provider's whim. Whatever capabilities they get, can be used "forever" going forward. And, at least flash can be ran "at home" with <10k in hardware, which isn't really possible / feasible with glm/k3 larger models.
The only movie I've ever enjoyed after reading the book was Jurassic Park. I've never seen the Dune movies or most other book-based movies, but honestly The Martian and Project Hail Mary were tempting. I'll probably surrender to one of those two someday.
The most recent Dune movies pull off the incredible feat of being a better telling of the tale than the book, but only if you already know the tale from the books. They’re not beginner-friendly, but they’re extremely respectful to the book reader, while still make conscious (and in my opinion, positive) changes to the story.
What are your goalposts? Depending on your requirements, there have been many moments of usable local AI. More recent ones were gpt-oss 120b and Qwen 3.6 27b.
I’ve heard that too, but since it’s a stealth model, I suspect they change it to whatever preview model provider that offers a free endpoint. In the past few months, I’ve gotten 4xx errors identifying the provider as DeepSeek.
distilling K3 into DS4 flash will likely only be a good idea for specialists. the difference in model capacity is otherwise too large.
we used to distill GLM 5.2 into Qwen 27B specialists with great success.
developing the workflows is tricky though. we had the advantage of a straightforward mapping function in mind: English -> SQL which was relatively easy to pipeline training for.
a specialist model sufficiently post-trained can outperform a frontier model while being dirt cheap.
what we did was distill GLM 5.2 into a 27B model on SQL and then post-train it with RL afterward. the result outperformed even Fable on that one task. the distillation step is just good sense in this workflow, to bootstrap a smaller model to the utmost you can before actually doing RL.
Do you have any reference guides on how to perform this distillation? I have a use case for this and hardware to run it on, but have never seen how it's done.
DS was serving the pro version at extremely low prices for a long time, and they've had integrations with opencode & other providers, so they likely gathered a lot of data from real developers doing real tasks (on openrouter they were labeled as such). Now they can use those live scenarios to further post-train their models and improve them further.
Can't wait to see if distilling k3 into dsv4 brings additional improvements. Anyway, having fast cheap models getting better is great for the community. Especially since these don't "go away" on a provider's whim. Whatever capabilities they get, can be used "forever" going forward. And, at least flash can be ran "at home" with <10k in hardware, which isn't really possible / feasible with glm/k3 larger models.