Comment by esafak
14 hours ago
It's open source; the price is up to the provider, and I do not see any on openrouter yet. ̶G̶i̶v̶e̶n̶ ̶t̶h̶a̶t̶ ̶d̶e̶v̶s̶t̶r̶a̶l̶ ̶i̶s̶ ̶m̶u̶c̶h̶ ̶s̶m̶a̶l̶l̶e̶r̶,̶ ̶I̶ ̶c̶a̶n̶ ̶n̶o̶t̶ ̶i̶m̶a̶g̶i̶n̶e̶ ̶i̶t̶ ̶w̶i̶l̶l̶ ̶b̶e̶ ̶m̶o̶r̶e̶ ̶e̶x̶p̶e̶n̶s̶i̶v̶e̶,̶ ̶l̶e̶t̶ ̶a̶l̶o̶n̶e̶ ̶5̶x̶.̶ ̶I̶f̶ ̶a̶n̶y̶t̶h̶i̶n̶g̶ ̶D̶e̶e̶p̶S̶e̶e̶k̶ ̶w̶i̶l̶l̶ ̶b̶e̶ ̶5̶x̶ ̶t̶h̶e̶ ̶c̶o̶s̶t̶.̶
edit: Mea culpa. I missed the active vs dense difference.
> Given that devstral is much smaller, I can not imagine it will be more expensive
Devstral 2 is 123B dense. Deepseek is 37B Active. It will be slower and more expensive to run inference on this than dsv3. Especially considering that dsv3.2 has some goodies that make inference at higher context be more effective than their previous gen.
Devstral is purely nonthinking too it’s very possible it uses less models (I don’t know how DS 3.2 nonthinking compares). It’s interesting because Qwen pretty much proved hybrid models work worse than fully separate models.
Deepseek v3.2 is that cheap because its attention mechanism is ridiculously efficient.
Yeah, DeepSeek Sparse Attention. Section 2: https://arxiv.org/abs/2512.02556