Thursday, October 29, 2026 🏢 AI Companies Hub RSS About Contact Admin
POPULAR BEATS: Generative AI LLMs & NLP Autonomous Agents Robotics & Hardware Enterprise AI AI Ethics & Policy 🏢 All AI Companies

DeepSeek V4 Flash 0731

Multi-head Latent Attention (MLA) + DeepSeekMoE
Official Source & Paper
DeepSeek logo DeepSeek 📄 Open Weights (MIT) / Cloud API 🌍 China (China) 🗓️ Effective Date: 31.07.2026

Ultra-efficient high-speed MoE model engineered with FP8 mixed-precision and Multi-token Prediction for unmatched cost efficiency.

Model Size / Parameters 238B (21B active) Architecture: Multi-head Latent Attention (MLA) + DeepSeekMoE
Context Window (Input) 256k tokens Token Capacity
Max Output Tokens 32k tokens Single Generation Limit
Supported Modalities Text, Code Native Multimodal Ingestion
License & Access Open Weights (MIT) / Cloud API Open Weights vs. API
Release Date July 2026 Effective Date: 31.07.2026

Verified Benchmark Results

Standardized academic and industry evaluation benchmarks (Effective Date: 31.07.2026).

MMLU-Pro / 5-Shot General Intellect
90.6%
SWE-bench Verified Software Engineering
61.2%
MATH 500 / AIME Mathematical Reasoning
93.1%
GPQA Diamond Doctoral Science
71.8%
HumanEval / Python Algorithm Synthesis
90.5%

Pricing & Cost Efficiency

Official API token pricing and self-hosted licensing terms.

Input Pricing (Prompt)
$0.07 / 1M tokens
Per 1 million processed tokens
Output Pricing (Generation)
$0.14 / 1M tokens
Per 1 million generated tokens
Verified Datasheet: Source: DeepSeek Research Documentation • Effective Date: 31.07.2026
← Back to All AI Model Datasheets

Latest Coverage & Analysis on DeepSeek V4 Flash 0731