AI Agent Hub
Back to models
DeepSeek: DeepSeek Pro Latest logo

DeepSeek: DeepSeek Pro Latest

Open Source ~deepseek Released 2026-08-13
36.0 / 100 1600.0B params 1M context Proprietary

About this model

DeepSeek Pro Latest is the rolling API product name for DeepSeek's flagship V4-Pro line (currently deepseek-v4-pro / DeepSeek-V4-Pro-0813). It is a sparse Mixture-of-Experts transformer with 1.6 trillion total parameters and about 49 billion activated parameters per token, trained on more than 32 trillion tokens and released under the MIT license with public weights on Hugging Face. The architecture combines Compressed Sparse Attention and Heavily Compressed Attention for million-token contexts, manifold-constrained hyper-connections, and Muon optimization, targeting strong reasoning, coding, and agentic workloads at lower inference cost than dense models of comparable capability.

The model supports non-thinking, Think High, and Think Max reasoning modes via OpenAI- and Anthropic-compatible APIs, with up to 1M input tokens and very long generated outputs (384K recommended for high/max reasoning). Official evaluations show leading open-weight results on graduate science (GPQA Diamond), hard knowledge suites (MMLU-Pro, HLE), competitive programming (LiveCodeBench, Codeforces rating), and software engineering benchmarks (SWE-Bench Verified), with substantial gains on agent benchmarks in the 0813 release (Terminal Bench 2.1, DeepSWE, Toolathlon-Verified).

DeepSeek Pro Latest is suited for chat, tool use, coding agents, long-document analysis, and research assistance. Self-hosting requires multi-GPU datacenter hardware; most teams consume it through DeepSeek's API or hosted inference providers, while open weights enable reproducibility and custom deployment with vLLM or SGLang including optional DSpark speculative decoding.

Benchmark Scores

BBH
87.5
HLE
42.7
DROP
88.7
MATH
64.5
MMLU
90.1
GSM8K
92.6
C-Eval
93.1
AGIEval
83.1
DeepSWE
62.7
CyberGym
83.3
MMLU-Pro
87.5
SimpleQA
57.9
GDPval-AA
1554.0
HellaSwag
88.0
HumanEval
76.8
BrowseComp
83.4
CodeForces
3206.0
GPQA-Diamond
90.1
LiveCodeBench
93.5
NL2Repo-Bench
61.5
SWE-Bench-Pro
55.4
AutomationBench
31.8
IMO-AnswerBench
89.8
Agents-Last-Exam
25.7
SWE-Bench-Verified
80.6
Terminal-Bench-2.0
67.9
Terminal-Bench-2.1
87.9
Toolathlon-Verified
74.1
SWE-Bench-Multilingual
76.2

Technical Specs

  • Parameters: 1600.0B
  • Architecture: MoE Transformer
  • Context Window: 1,000,000 tokens
  • Input Modalities: text

Hardware Requirements

  • VRAM: 960.0 GB
  • Compute: 8x NVIDIA B200 192GB (single-node EP/DP; ~960 GB aggregate GPU memory for FP4+FP8 weights)

Pricing

Input Output Currency
0.18 / 1M tokens 8.00 / 1M tokens USD