DeepSeek V4 Launch: Open‑Source Model Beats Closed‑Source Leaders in Coding & Math, 1.6 T Params, 1 M Context

Model variants

DeepSeek‑V4‑Pro – 1.6 T total parameters, 49 B activated, FP4+FP8 precision, 1 M token context window.

DeepSeek‑V4‑Flash – 284 B total parameters, 13 B activated, same precision and context.

Benchmark performance

Programming ability

LiveCodeBench: V4‑Pro 93.5, Opus 4.6 88.8, Gemini‑3.1‑Pro 91.7.

Codeforces competition score: V4‑Pro 3206, GPT‑5.4 3168, Gemini‑3.1‑Pro 3052.

SWE‑bench verified: V4‑Pro 80.6, Opus 4.6 80.8, Gemini‑3.1‑Pro 80.6.

Mathematical reasoning

HMMT 2026: V4‑Pro 95.2 %, Opus 4.6 96.2 %, GPT‑5.4 97.7 %.

IMO QA: V4‑Pro 89.8 %, Opus 4.6 75.3 %, GPT‑5.4 81.0 %, Gemini‑3.1‑Pro 86.0 %.

Apex Shortlist high‑difficulty: V4‑Pro 90.2 %, Opus 4.6 85.9 %, GPT‑5.4 78.1 %, Gemini‑3.1‑Pro 89.1 %.

Chinese knowledge benchmarks

MMLU‑Pro: V4‑Pro 87.5, Opus 4.6 87.1, GPT‑5.4 87.5, Gemini‑3.1‑Pro 91.0.

Chinese‑SimpleQA: V4‑Pro 84.4, Opus 4.6 76.4, GPT‑5.4 76.8, Gemini‑3.1‑Pro 85.9.

SimpleQA‑Verified: V4‑Pro 57.9, Opus 4.6 46.2, GPT‑5.4 45.3, Gemini‑3.1‑Pro 75.6.

Agent capability

Terminal Bench 2.0: V4‑Pro 67.9, Opus 4.6 75.1, GPT‑5.4 68.5, Gemini‑3.1‑Pro 66.7.

MCPAtlas tool calls: V4‑Pro 73.6, Opus 4.6 73.8, GPT‑5.4 67.2, Gemini‑3.1‑Pro 66.6.

Toolathlon usage: V4‑Pro 51.8, Opus 4.6 47.2, GPT‑5.4 54.6, Gemini‑3.1‑Pro 50.0.

Million‑token context

One million tokens correspond to roughly 15–20 full novels or the entire source code of a medium‑size project. Many models exhibit "context decay" in the latter half of long inputs. V4 addresses this by redesigning the attention mechanism.

Mixed attention architecture

V4 Attention Mechanism:
├─ CSA (Compressed Sparse Attention) – compresses on token dimension
└─ HCA (Heavy Compression Attention) – drastically reduces KV‑cache consumption

Efficiency comparison

Single‑token inference FLOPs: baseline V3.2 100 %, V4‑Pro 27 % (73 % reduction).

KV‑cache consumption: baseline V3.2 100 %, V4‑Pro 10 % (90 % reduction).

Processing one million tokens therefore requires about one‑quarter the compute and one‑tenth the VRAM of V3.2.

Inference modes

Non‑think : fast response for everyday tasks; output is direct.

Think High : logical analysis for complex or planning tasks; returns thought process plus summary.

Think Max : extreme reasoning to explore model limits; requires special prompt and returns detailed reasoning.

API usage example

# Official API
from openai import OpenAI

client = OpenAI(
    api_key="your-api-key",
    base_url="https://api.deepseek.com"
)

# V4‑Pro
response = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[{"role": "user", "content": "Hello"}]
)

# V4‑Flash
response = client.chat.completions.create(
    model="deepseek-v4-flash",
    messages=[{"role": "user", "content": "Hello"}]
)

Model download links

V4‑Pro – huggingface.co/deepseek-ai/DeepSeek-V4-Pro (ModelScope: modelscope.cn/models/deepseek-ai/DeepSeek-V4-Pro)

V4‑Flash – huggingface.co/deepseek-ai/DeepSeek-V4-Flash (ModelScope: modelscope.cn/models/deepseek-ai/DeepSeek-V4-Flash)

V4‑Pro‑Base – huggingface.co/deepseek-ai/DeepSeek-V4-Pro-Base (ModelScope: modelscope.cn/models/deepseek-ai/DeepSeek-V4-Pro-Base)

V4‑Flash‑Base – huggingface.co/deepseek-ai/DeepSeek-V4-Flash-Base (ModelScope: modelscope.cn/models/deepseek-ai/DeepSeek-V4-Flash-Base)