April 2, 2026QUARANTINEdeepseek· 5 min read

DeepSeek-V3: a 671B open MoE, and what tiering one involves

DeepSeek-V3 is a 671-billion-parameter mixture-of-experts that activates 37B parameters per token, trained on 14.8 trillion tokens and shipped natively in FP8. It holds the Quarantine tier in our registry. It's also a useful case for getting the arithmetic right — and for explaining what our tiers actually mean.

Updated August 11, 2026 — this post originally reported the Hardened tier and described a witness-cosign model that no longer exists. Hardened now requires that we downloaded and hash-checked the weights ourselves, which at 671 GB we have not done. The tier stated here is the one the registry actually shows.

tldr

671B total / 37B active MoE, 128K context, native FP8 weights (~671 GB on disk; the HF repo lists ~685B params including the multi-token-prediction module). Code is MIT; the weights carry the DeepSeek Model License (commercial use permitted). A common error: calling the ~671 GB figure "FP16." It isn't — that's the FP8 footprint. At BF16 the weights would be ~1.3 TB. Quarantine in our registry; verify on the live artifact page.

parameters
671 B (MoE, 37 B active)
training tokens
~14.8 T
context
128K tok
native precision
FP8
license
MIT (code) + DeepSeek (weights)
on-disk
~671 GB FP8 · ~1.3 TB BF16
tier (registry)
QUARANTINE
released
2024-12-26

Get the sizing right

DeepSeek-V3 was trained and released natively in FP8 (E4M3). At one byte per parameter, 671B parameters is roughly 671 GB on disk — the Hugging Face repo reports ~685B parameters including the multi-token-prediction module, hence the ~685 GB figures you'll see. The number people most often get wrong is the precision label: 671 GB is the FP8 footprint. If you actually wanted BF16 weights, that's two bytes per parameter — about 1.3 TB. Halving FP8 to "336 GB" or calling FP8 "FP16" are both off by a factor of two; the registry and this post use the FP8-native figure.

What the tiers mean

A blog that only announces good news is propaganda. The value of the catalog is that the tier reflects evidence, not enthusiasm:

  • Closed — a third party reports the artifact exists and we recorded their numbers. We hold no weights and no publisher hash; every figure is someone else's claim.
  • Quarantine — we hold the publisher's own hash, logged and signed, but have not yet downloaded the weights to check it ourselves. DeepSeek-V3 sits here — at 671 GB, pulling it is not a trivial step. Pull it for research; tier-gated production policies won't route to it.
  • Hardened — we downloaded the weights, verified their SHA-256 against the publisher's hash, and ran a supply-chain scan mapped to NIST AI RMF / COSAIS, all signed by our own key. This is the only tier where the evidence is ours.
  • Sealed — was to be independent third-party domain review. It depended on external witness cosigns that were never genuinely third-party, so the tier is retired and no artifact holds it. We would rather show you a gap than a badge we can't back.

You don't have to trust our label. Open the live artifact page for its current tier and provenance, or verify any page at /verify.

Sources

realistic deployment footprint

Precision
Weights
Runtime
Fits from
FP8 (native)
native
~671 GB
vLLM / SGLang, multi-node
multi-GPU
FP8
~671 GB + KV
8× H200 141GB or 2× 8×H100
BF16
—
~1.3 TB
multi-node only

more releases