Open weights, from 2.3B to 2.8 trillion.
The open frontier stopped being one category. Seven models on this page span three orders of magnitude: one runs on a laptop, one does not fit on an eight-GPU node at all, and the interesting ones sit in between. This is what each actually costs to serve on Blackwell, with every number linked to where it came from.
The table
sourced, not estimatedAnything marked derived is arithmetic I did myself because the vendor publishes the format but not the byte count. Hover it for the working. Everything else is lifted from a primary source and linked below the card. None of it was measured on my own hardware, which bothers me enough that I wrote down the harness I would run to fix it.
Will it fit?
pick your hardwareThe models
what each one is betting onThe drafter
not a model you hostWhat I take from this
opinion, clearly labelledActive parameters won
Total parameters on this page span a factor of ninety, from Gemma 4's largest dense variant to Kimi K3. Active parameters span a factor of eight, from 13B to 104B. That compression is the whole story of the last year: models got enormously larger and only modestly more expensive per token. Total parameters decide whether you can host it. Active parameters decide what it costs you on every token, forever.
The KV cache is the product
GLM-5.3-Flash's IndexPool gets 4.44x on cache size. DeepSeek V4 reaches 10% of V3.2's cache at 1M context. Neither of those is a footnote. A million-token context is not an attention-math achievement, it is a memory-footprint achievement, and the attention redesign is downstream of that.
Read the licence before the benchmark
GLM-5.3-Flash is MIT. GLM-5.3 is not, it ships under a bespoke licence with the same vendor's name on it. That distinction will cost some team a deployment, and it is one line in a model card that most benchmark roundups do not print.
The range is the finding
A year ago this page would have been one category. It is now three: models that run on a laptop, models that take part of a node, and models that do not fit on a node at all. Kimi K3 at 1.56 TB forces tensor parallelism across machines and puts the all-reduce on a NIC; Gemma 4's small variants do not need a GPU cluster to exist. Those are not the same engineering problem and it is a mistake to plan for them with one playbook.
Speculation stopped being optional
DeepSeek ships MTP and DSpark configs in the recipe. DFlash 2 gets 2.7-4.6x losslessly. When the drafter is in the published serving config rather than a paper appendix, speculative decoding has stopped being a research result and become part of the runtime.