First impressions continue to improve with Step-3.7-flash. I built one of the Northwind full CRUD database example apps with it, and the result compared well with the versions by Deepseek-4-pro, Qwen 3.6, and other favorite models. It only took 2 prompts to complete with step-3.7-flash, and would likely have required only 1 prompt if I'd asked for full CRUD functionality in the initial query:

https://com-pute.com/nick/northwind-flask-app-stepfun-flash-3.7-no-db.zip

Along the way, the model provided a link to a web site containing the Northwind database as a downloadable SQLite file, and built the code of the app to work with this file:

https://github.com/jpwhite3/northwind-SQLite3

Results worked perfectly first shot.

Then I had the agent build a zip file package with everything needed to install the app on another server, and it yielded the file linked above, which included detailed install instructions.

I also had it build a 3D space invaders game. This was the first-shot result - bullets don't work yet, but it's a decent initial output:

https://com-pute.com/nick/space-invaders-3d-step37flash-1.html

Additionally, I asked the model set up Rustdesk to be restarted every time the system it was on restarted, and it did a perfect job first shot (Pi was used as the harness).

So far, step-3.7-flash feels more similar to a frontier model than most of the other models I've self hosted. I do wish it was just a bit smaller at q4 compression, because there's not a ton of room for KV cache (I had it set up to use ~89k, and did run into it just running out of cache for a long task).

I don't think step-3.7-flash will dethrone Qwen 3.6 for basic coding tasks yet, but it's another great tool - the best yet for knowledge, which still runs at a fast speed.