DeepSeek V4

Name: DeepSeek V4
Author: DeepSeek

DeepSeek

Open-weight model · runs on-premises · Published on 5 July 2026 · Updated on 18 October 2026

DeepSeek V4 brings together two mixture-of-experts models released under the MIT licence, with a native context of 1,048,576 tokens. Since 17 October 2026 the deepseek-flash API name has designated DeepSeek V4.1 Flash 1017, a multimodal model whose backbone totals 552 billion parameters, while the Pro 0813 version remains published at 1.6 trillion.

Key points: Price-to-performance ratio · Mixture of experts · Flash 1017 552B, 16B active when decoding · Pro 0813 1.6T, 49B active · Multimodal on the Flash side · Compatible with vLLM and ninfer.

Architecture

DeepSeek V4.1 Flash 1017 totals 552 billion backbone parameters but activates only 16 billion per token when decoding, then just 8 billion during prefill, a gap explained by its causal encoder-decoder architecture. It spreads 384 routed experts and one shared expert across forty layers, and weighs 510 GB published in FP8 with FP4 experts.

Its main contribution concerns the attention cache, since second-generation compressed sparse attention reuses projections across layers and keeps the cache in FP4. The persistent footprint in video memory therefore falls to 890 bytes per token, roughly a quarter of that of V4 Flash.

DeepSeek V4 Pro 0813 totals 1.6 trillion parameters, 49 billion of them active per token, for 865 GB of weights published in mixed FP4 and FP8, with no image handling. Its API name has routed to the Flash since 14 September 2026, pending a V4.1 Pro that does not exist to date.

DeepSeek V4 Flash 0731, which totalled 284 billion parameters including 13 billion active for 167 GB published, left the API on 10 September 2026, and its name now serves the 1017 at the Flash rate. The weights remain downloadable, so the guides in this notebook that size it keep their value for that precise revision.

Strengths

On the benchmarks published by the vendor, the 1017 gains 11.1 points on Terminal-Bench 4.0, at 42.3% against 31.2%, and moves from 74.2% to 77.0% on DeepSWE v1.1, where it overtakes Claude Opus 5. It nonetheless remains behind Opus 5 on the last-resort examination, at 42.6% against 56.3%.

The API price has not moved with this update, since it remains $0.15 per million input tokens and $0.60 on output during off-peak hours, double that at peak hours, and $0.003 on a cache hit. The vendor also announces shorter answers for an equivalent task.

Use cases

The Pro version handles demanding reasoning tasks, while the Flash serves everyday work with speed, interprets images and targets development agents in particular, where the input volume largely dominates the output. That asymmetry justifies a prefill cost halved.

Deployment

Served by vLLM or llama.cpp depending on quantisation, or by ninfer on a single RTX 5090 or RTX PRO 6000 GPU.

All models on the platform are interchangeable through a single gateway: switching models is a matter of changing one configuration line, with no code rewrite. Local execution remains the priority.

Official sources

Which machine can run DeepSeek V4?

DeepSeek V4 Pro totals 1600 billion parameters, about 920 GB in NVFP4 (calculated as 1,600 billion × 0.5 byte plus a 15% execution margin). 4 platforms in the catalogue have enough memory for the Pro version. Active parameters govern speed, not footprint: every expert stays resident. DeepSeek V4 Flash, 284 billion parameters, weighs about 163 GB by the same calculation (167 GB of weights published in FP4 + FP8) and also fits on a DGX Station GB300 and on a cluster of two DGX Spark, see the DeepSeek V4 Flash on DGX guide.

Platform	Memory	Most precise format that fits	Sizing page
H200 SXM server	1,128 GB	NVFP4	DeepSeek V4 on H200 SXM server
B200 SXM	1,440 GB	NVFP4	DeepSeek V4 on B200 SXM
B300 SXM	2,304 GB	FP8	DeepSeek V4 on B300 SXM
GB300 NVL72	20,700 GB	FP16	DeepSeek V4 on GB300 NVL72

See also DeepSeek V4 in NVFP4 and the full sizing matrix.