model
DeepSeek-V3 is a mixture-of-experts large language model with 671B total parameters but only 37B activated per token via top-k gating.
No evidence-backed relationships are recorded.
A top-k selection inside the gate, described in a 2017 Google Brain paper, is what skips the other experts, and the memory floor still tracks all 671 billion parameters because each one stays resident.
Reality
A JAX and Transformer Engine version of the same GB200 run reports 1,068 TFLOPS per GPU, a 10.4x gain. The 84% share bounds what fixing communication alone could have bought, so the expert matmuls got faster too.
Reality
The joint NSA, CISA and FBI advisory names six China-based companies and describes billions of tokens taken through paid APIs, cloud resellers and proxy transfer stations, which turns billing telemetry into a detection surface.
Reality
The study counts only complexity and dead code, because those are the two pyscn metrics that stay exact when you analyze just the files a patch touched. The human's own commit trips the same rule 24% of the time.
Reality
The three-model set topped out at a US plus China pairing, so Mistral Small 3.2 went in late; the resulting China plus EU pair posted the run's best score and its worst capitulation rate, for about $0.15 in stranded reviews.
Reality
Intern-S2-Preview-397B puts page-level paper reading and tool use into Apache 2.0 weights. The download listing says 810 GB, and every benchmark number so far is the lab's own.
Reality
A conversion-pipeline checklist item, "MTP round-trip", turns on a distinction teams collapse: the training-time auxiliary loss is disposable, the inference-time draft head is not.
Reality