DeepSeek V4 Pro: 1.6 trillion parameters, 805 GiB on disk

A forwarded post about DeepSeek V4 quotes round benchmark figures and says a single machine is enough for the large model. We do not know where it came from, so we do not link it. Instead we read the public model cards and measured the weights ourselves.

DeepSeek V4 Pro is an open mixture-of-experts language model from DeepSeek with 1.6 trillion parameters, 49 billion of them active per token, and a context window of one million tokens (model card, MIT licence). The experts are stored in FP4 and most of the rest in FP8. That is why the weights take 805 GiB rather than the roughly 3,000 GiB that 16 bits per parameter would give.

DeepSeek V4 is not news this week. The DeepSeek-V4-Pro repository was created on 22 April 2026 and its card was last modified on 22 June 2026 (Hugging Face API, queried 5 November 2026). Further repositories followed: Flash-0731 on 31 July, Pro-0813 on 13 August, V4.1-Flash on 10 September and V4.1-Flash-1017 on 17 October. How the newest one behaves is covered in our post DeepSeek 1017: old model name, new weights.

Correction of 5 November 2026: we re-read the 30 October version against the sources. We changed the byte total of the shards (now 864,721,029,744), the arithmetic values in the chart, four rows of the benchmark table, the price row for cached input and the statements about our gateway. We had misfiled the figures 87.1 and 88.3.

What sits on the disk

The card states 1.6 trillion parameters with 49 billion active, and the precision "FP4 + FP8 Mixed": expert parameters are FP4, most others FP8. The config.json lists 384 routed experts with 6 active per token.

On 5 November 2026 we queried the file listing through the Hugging Face API with ?blobs=true and added up the sizes of the 64 .safetensors files. The total is 864,721,029,744 bytes, or 805.3 GiB. The API metadata counts 1,598,839,674,782 parameters. The arithmetic in the chart follows from that.

The 805.3 GiB are 6.3 times the memory of the four cards, and the context cache comes on top. Anyone deriving three terabytes from 1.6 trillion parameters is too high, yet the model stays in data-centre territory. The active parameter count changes nothing here: all 384 experts must be available, because every token can pick different ones.

What the benchmark figures say in the cards

We compared the figures from the forwarded post with four cards: DeepSeek-V4-Pro (22 June 2026), Pro-0813, Flash-0731 and V4.1-Flash. Every value in the right-hand column is a DeepSeek claim. We did not run any of these tests.

So several of the post's numbers match real cells, only in another row, another mode or another model. We cannot say how they travelled. Anyone passing on a number needs to name the row, the mode and the card date with it.

What the card says about the architecture

These are also vendor statements, from the card and the technical report (arXiv 2606.19348):

• Two attention methods in alternation, Compressed Sparse Attention and Heavily Compressed Attention. At a one-million-token context, the card says Pro needs 27% of the per-token compute and 10% of the KV cache of DeepSeek-V3.2.
• Manifold-Constrained Hyper-Connections, meant to keep the signal stable across layers.
• The Muon optimiser and pre-training on more than 32 trillion tokens.

According to its card, Pro-0813 is the official release that supersedes the preview, with a speculative decoding module (DSpark) attached. As a serving example the card gives a single node with four GB300.

API prices

We fetched DeepSeek's price page on 5 November 2026 from a server in our house (dsh), because the TLS connection from the workstation PC is reset. Prices are in US dollars per 1 million tokens. Peak hours are Monday to Friday 01:00 to 04:00 and 06:00 to 10:00 UTC, except Chinese public holidays; off-peak is half price.

The page lists the version DeepSeek-V4-Pro-0813 for deepseek-v4-pro and DeepSeek-V4.1-Flash-1017 for deepseek-flash. According to its footnote, the legacy Flash names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but the models behind them are retired. V4.1-Flash-1017 serves those requests at the Flash price. Prices quoted from before 5 November therefore say little about today's bill.

On Pro a cache hit costs one thirtieth of a miss (0.022 against 0.66), on Flash one fiftieth (0.003 against 0.15). How much of our own traffic comes from the cache is measured in vLLM prefix caching in self-hosting.

What this means for us

Pro stays out of reach for us: 805.3 GiB of weights against 127.4 GiB of card memory. For customer data only a locally run model is an option for us. Our assessment of where DeepSeek's API is hosted, and of the data-protection question, is in AI radar: DeepSeek, from R1 to the V4 series.

For a purchase of our own, the smaller variant is the yardstick. How far the V4-Flash class gets on pooled memory we measured in two DGX Spark in our own test. More assessments of new models are collected in AI explained: models, market, news. If you want to check which model size fits the hardware you already own, our approach is described under Corporate LLM in-house.

Further Reading

• Model card DeepSeek-V4-Pro (size statements, benchmark table, as of 22 Jun 2026): huggingface.co/deepseek-ai/DeepSeek-V4-Pro
• File listing with shard sizes, basis of the 805.3 GiB: Hugging Face API with blobs=true
• Model card DeepSeek-V4-Pro-0813 (Terminal Bench 2.1, serving example): huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813
• Model card DeepSeek-V4-Flash-0731 (table of 31 Jul 2026): huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731
• Technical report on DeepSeek-V4: arxiv.org/abs/2606.19348
• DeepSeek API price list, fetched 5 Nov 2026: api-docs.deepseek.com