DeepSeek-V4 collection release (Flash + Pro, up to 1.6T)

TL;DR

DeepSeek-AI publishes the DeepSeek-V4 collection on Hugging Face: a four-checkpoint family spanning two sizes (Flash and Pro) and two post-training stages (Base and chat-tuned) — DeepSeek-V4-Flash-Base (292B), DeepSeek-V4-Flash (158B), DeepSeek-V4-Pro-Base (1.6T), and DeepSeek-V4-Pro (862B). All four were updated within the past month and the chat variants together account for ≈7.5M Hub downloads in the first ~18 days, with the Pro chat variant alone at ≈4.67M. Pro-Base is the largest publicly released open-weights checkpoint to date at 1.6T parameters; the Pro chat checkpoint at 862B is presumably the deployment-shaped (active or post-pruned) variant. The HF collection page itself is the only artifact at filing time — no technical report, model cards, or benchmark numbers are linked from the collection landing.

Key claims

The collection contains four checkpoints organized as a 2×2 grid of {Flash, Pro} × {Base, chat} [collection listing].
Reported parameter counts on the HF cards: Flash-Base 292B, Flash 158B, Pro-Base 1.6T, Pro 862B [collection listing].
The Base checkpoints (292B, 1.6T) and the chat checkpoints (158B, 862B) within each size differ in parameter count, indicating either post-training pruning/compression of the Base or that the Base reports total MoE parameters while the chat card reports the deployed/active variant — the collection page does not disclose which [collection listing].
The chat variants are tagged “Text Generation” and the Pro-chat variant has ≈4.67M downloads in roughly 18 days; the Flash-chat variant has ≈2.84M [collection listing].
The collection has been upvoted 655 times on Hugging Face [collection header].

Method

The HF collection is a release announcement only — no architecture, training recipe, dataset description, or benchmark table is attached at the collection level, and the linked model cards were not fetched at filing time. Inferences from the file shape and DeepSeek’s prior releases (V3, V3.1, V3.2, R1 — all visible in the same deepseek-ai collections sidebar): the V-series has consistently been DeepSeek-MoE-style sparse mixture-of-experts with MLA attention, FP8 mixed-precision pretraining, and large total-parameter / small activated-parameter ratios; V3 was 671B-total / 37B-active. The 1.6T total of Pro-Base, paired with an 862B chat variant, is consistent either with a sparse-MoE base where the chat variant has been pruned/distilled, or with a base-vs-deployment naming convention where the larger number is total experts and the smaller is something closer to deployable/active scale. Confirmation requires reading the individual model cards.

Results

No benchmarks at the collection level. Adoption signal: the chat checkpoints together have served ≈7.5M downloads in ≈18 days post-release [collection listing], which — for an open-weights LLM at this size class — is in the same range as DeepSeek’s earlier V3/V3.1 releases at comparable post-release intervals.

Why it’s interesting

If the 1.6T number on DeepSeek-V4-Pro-Base is the total parameter count of an open-weights checkpoint actually uploaded to HF, this is the largest open-weights base model published to date — past both arcee-ai/Trinity-Large-Base (400B total) and tencent/Hunyuan-A13B-derived tencent/Hy3-pr (295B total). For Open foundation-model releases this extends the upper-end datapoint of the catalog by ~4×. The 2×2 release shape (size × {Base, chat}) is the same convention used by Trinity Large (Trinity Large: An Open 400B Sparse MoE Model ships TrueBase / Base / Preview) and HunyuanImage 3.0 (HunyuanImage 3.0 Technical Report ships base + Instruct), and the Flash/Pro tier split mirrors the deployment-tiering convention OpenAI/Anthropic/Google use for closed models. The interesting open question is whether the Base→chat parameter drop (292B→158B Flash, 1.6T→862B Pro) reflects expert pruning, distillation into a smaller dense (or less sparse) backbone, or simply a “total vs active” naming inconsistency on the HF cards — the collection page does not say.