DeepSeek-V4-Pro

How it is built

At 1.6T parameters, DeepSeek-V4-Pro is the largest model DeepSeek has released. A token runs through 49B of them.¹¹1DeepSeek, DeepSeek-V4-Pro (preview) model card, 24 April 2026. FLOPs and KV cache compared with DeepSeek-V3.2 at a 1M-token context; "MoE expert parameters use FP4 precision; most other parameters use FP8". https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro ²²2DeepSeek, DeepSeek-V4.1-Flash model card, 10 September 2026. Its base-model table lists V4-Flash at 284B, V4-Pro at 1.6T and V4.1-Flash at 552B. Agentic benchmarks at maximum reasoning effort; SimpleQA-Verified is the 25-shot exact match of the base models. https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash

Every one of the 61 layers carries 384 routed experts and 1 shared expert. A token uses 6 of the routed ones.³³3DeepSeek, DeepSeek-V4-Pro-0813 model card and config.json (num_hidden_layers 61, n_routed_experts 384, n_shared_experts 1, num_experts_per_tok 6, num_hash_layers 3). Benchmarks in the minimal mode of DeepSeek Harness, max reasoning effort, temperature 1.0, top_p 0.95. https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813 In the first 3 layers a hash of the token ID fixes the experts.⁴⁴4DeepSeek, DeepSeek-V4 technical report, arXiv 2606.19348: the first dense layers replaced by hash-routed MoE layers, and affinity scores computed with Sqrt(Softplus) instead of Sigmoid. https://arxiv.org/abs/2606.19348 The later layers learn their routing and score experts with Sqrt(Softplus), where DeepSeek-V3 used a sigmoid.⁴⁴

Attention combines Compressed Sparse Attention (CSA) with Heavily Compressed Attention (HCA). At a 1M context DeepSeek puts the model at 27% of the single-token inference FLOPs of DeepSeek-V3.2 and 10% of its KV cache.¹¹ Manifold-Constrained Hyper-Connections (mHC) reinforce the residual connections.¹¹ The routed experts are stored in FP4, most other weights in FP8.¹¹

What is new

The first version, published on 24 April 2026, was a preview with open weights.⁵⁵5DeepSeek API Docs, DeepSeek-V4 preview release note, 24 April 2026. https://api-docs.deepseek.com/news/news260424 DeepSeek-V4-Pro-0813 is the official release. It keeps the preview's structure and adds a DSpark module for speculative decoding.³³

The preview scored 12.8 on DeepSWE. The 0813 snapshot scores 62.7.³³ On Terminal Bench 2.1 it moved from 72.1 to 87.9, in the minimal mode of DeepSeek Harness at max effort.³³ Reasoning effort now has three levels, from low to max.³³

In September DeepSeek released the smaller DeepSeek-V4.1-Flash. Its card has it ahead of V4-Pro on every agentic benchmark the two share.²² V4-Pro keeps the lead on knowledge: its base model scores 55.2 on SimpleQA-Verified, against 42.3.²²

Running it

The 0813 weights carry the MIT licence.³³ The card's vLLM example serves them on a single node of four GB300 GPUs, and one flag turns DSpark on.³³ At high and max effort it recommends room for 384K output tokens.³³

The API is less settled. The pricing page maps deepseek-v4-pro to DeepSeek-V4-Pro-0813, with 1M tokens of context and 384K of output. Images are not accepted.⁶⁶6DeepSeek API Docs, Models and Pricing, as listed on 29 September 2026. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, excluding Chinese public holidays. https://api-docs.deepseek.com/quick_start/pricing At peak a million input tokens cost $1.32, or $0.044 on a cache hit, and a million output tokens $3.96. Off-peak prices are half.⁶⁶

The release note of 10 September contradicts it. From 14 September, it states, every deepseek-v4-pro request goes to V4.1-Flash at V4.1-Flash prices, until a V4.1-Pro launches.⁷⁷7DeepSeek API Docs, DeepSeek-V4.1-Flash release note, 10 September 2026. The routing starts on 14 September at 04:00 UTC. https://api-docs.deepseek.com/news/news260910 The pricing page does not mention that routing.⁶⁶