Post History

I've continued to compare coding output with Qwen 3.6 models, and Qwen is just higher quality, faster, and more reliable that step-3.7-flash (not to mention, the MOE versions of qwen 3.6 can run on very small consumer GPUs).

I've also done more comparisons with the Qwen 3.5 122b model, and its knowledge depth and speed are in the same ballpark as that of step-3.7-flash. I think the bigger Qwen model is better than step-3.7-flash at coding and agentic tasks.

If/when we get a Qwen 3.7 model in that size ballpark, I expect it will be king for the class of machines that include DGX Spark and Strix Halo.

In the meantime, step-3.7-flash is very capable, and it will be a great addition to the stable, for those situations when the perspective of another model is valuable.

I've continued to compare coding output with Qwen 3.6 models, and Qwen is just higher quality, faster, and more reliable (not to mention, it can run on very small consumer GPUs).

I've also done more comparisons with the Qwen 3.5 122b model, and its knowledge depth and speed are in the same ballpark as that of step-3.7-flash. I think that bigger Qwen model is better than step-3.7-flash at coding and agentic tasks. If/when we get a Qwen 3.7 model in that size ballpark, I expect it will be king for the class of machines that include DGX Spark and Strix Halo.

In the meantime, step-3.7-flash is very capable, and I think will be a great addition to the stable, for those situations when the perspective of another model is valuable.

Previous Versions