What appeared
NVIDIA dates Dory to October 16. This English-language model has four billion parameters and reuses shared layers to vary computation without another checkpoint.
The trade-off is more than model size
NVIDIA’s one-, two-, and four-loop results improve unevenly. These are provider measurements, not our tests; quantized versions may differ.
An optional cache setting limits memory growth across loops, not total memory or changed-answer risk.
Check the software before the download
At our October 18 source check, the Dory llama.cpp change was still a draft pull request and the vLLM change was open. The model card points to development builds rather than established releases.
The llama.cpp proposal lists incomplete long-context qualification and untested Metal support. A model-card setup example is not a guarantee that every desktop app, operating system, or hardware combination supports it.
Our take
For AI geeks and experienced technicians, the useful lesson is to separate stored model size from the work needed to answer. Parameter count alone cannot tell you how fast an app will feel.
Before choosing a trial, confirm the exact runtime supports your format and settings. Use checkable tasks and compare answer quality, peak memory, and elapsed time together. We would not choose a winner from one advertised score. We have not downloaded, run, or benchmarked Dory for this report.