# Added benchmarks are from the original V4 preview instruct-model comparison, not the later 0731/0813 releases.
name = "DeepSeek V4 Pro"
description = "Open MoE flagship with million-token context for coding and long agent runs"
family = "deepseek-thinking"
release_date = "2026-04-24"
last_updated = "2026-04-24"
attachment = false
reasoning = true
temperature = true
tool_call = true
structured_output = true
knowledge = "2025-05"
open_weights = true

[limit]
context = 1_000_000
output = 384_000

[modalities]
input = ["text"]
output = ["text"]

[[weights]]
label = "Hugging Face"
url = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"

[[benchmarks]]
name = "SWE-Bench Verified"
score = 80.6
metric = "resolved"
source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"

[[benchmarks]]
name = "Artificial Analysis Coding Agent Index"
score = 50.1
metric = "average pass@1"
harness = "Claude Code"
variant = "high"
source = "https://artificialanalysis.ai/agents/coding-agents"

[[benchmarks]]
name = "SWE-Atlas Codebase QnA"
score = 67.8
metric = "pass@1"
harness = "Claude Code"
variant = "high"
source = "https://artificialanalysis.ai/agents/coding-agents"

[[benchmarks]]
name = "SWE-Bench Pro"
score = 18
metric = "pass@1"
harness = "Claude Code"
variant = "high"
dataset = "hard-aa"
source = "https://artificialanalysis.ai/agents/coding-agents"

[[benchmarks]]
name = "Terminal-Bench"
score = 64.7
metric = "pass@1"
harness = "Claude Code"
variant = "high"
version = "2.1"
source = "https://artificialanalysis.ai/agents/coding-agents"

[[benchmarks]]
name = "MMLU-Pro"
score = 87.5
metric = "EM"
variant = "preview checkpoint; max effort"
source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"

[[benchmarks]]
name = "SimpleQA-Verified"
score = 57.9
metric = "pass@1"
variant = "preview checkpoint; max effort"
source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"

[[benchmarks]]
name = "Chinese SimpleQA"
score = 84.4
metric = "pass@1"
variant = "preview checkpoint; max effort"
source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"

[[benchmarks]]
name = "GPQA Diamond"
score = 90.1
metric = "pass@1"
variant = "preview checkpoint; max effort"
source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"

[[benchmarks]]
name = "Humanity's Last Exam"
score = 37.7
metric = "pass@1"
variant = "preview checkpoint; max effort; without tools"
source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"

[[benchmarks]]
name = "LiveCodeBench"
score = 93.5
metric = "pass@1"
variant = "preview checkpoint; max effort"
source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"

[[benchmarks]]
name = "Codeforces"
score = 3206
metric = "rating"
variant = "preview checkpoint; max effort"
source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"

[[benchmarks]]
name = "HMMT"
score = 95.2
metric = "pass@1"
variant = "preview checkpoint; max effort"
dataset = "February 2026"
source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"

[[benchmarks]]
name = "IMOAnswerBench"
score = 89.8
metric = "pass@1"
variant = "preview checkpoint; max effort"
source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"

[[benchmarks]]
name = "MathArena Apex"
score = 38.3
metric = "pass@1"
variant = "preview checkpoint; max effort"
source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"

[[benchmarks]]
name = "MathArena Apex Shortlist"
score = 90.2
metric = "pass@1"
variant = "preview checkpoint; max effort"
source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"

[[benchmarks]]
name = "MRCR"
score = 83.5
metric = "MMR"
variant = "preview checkpoint; max effort"
dataset = "1M context"
source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"

[[benchmarks]]
name = "CorpusQA"
score = 62
metric = "accuracy"
variant = "preview checkpoint; max effort"
dataset = "1M context"
source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"

[[benchmarks]]
name = "Terminal-Bench"
score = 67.9
metric = "accuracy"
variant = "preview checkpoint; max effort"
version = "2.0"
source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"

[[benchmarks]]
name = "SWE-Bench Pro"
score = 55.4
metric = "resolved"
variant = "preview checkpoint; max effort"
source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"

[[benchmarks]]
name = "SWE-Bench Multilingual"
score = 76.2
metric = "resolved"
variant = "preview checkpoint; max effort"
source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"

[[benchmarks]]
name = "BrowseComp"
score = 83.4
metric = "pass@1"
variant = "preview checkpoint; max effort"
source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"

[[benchmarks]]
name = "Humanity's Last Exam"
score = 48.2
metric = "pass@1"
variant = "preview checkpoint; max effort; with tools"
source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"

[[benchmarks]]
name = "MCP Atlas"
score = 73.6
metric = "pass@1"
variant = "preview checkpoint; max effort"
dataset = "public"
source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"

[[benchmarks]]
name = "GDPval-AA"
score = 1554
metric = "Elo"
variant = "preview checkpoint; max effort"
source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"

[[benchmarks]]
name = "Toolathlon"
score = 51.8
metric = "pass@1"
variant = "preview checkpoint; max effort"
source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"