DeepSeekV4
#76 en Modèles de langage open sourceDeepSeek · v4 · depuis 24. April 2026 (Preview); V4-Flash GA: 31. Juli 2026; V4-Pro GA: 12./13. August 2026 · 2× · vu le 18 août 2026
17
Momentum
DeepSeek V4 is an open-source language model family from DeepSeek, comprising the V4-Pro variant (1.6 trillion total parameters, 49B active) and V4-Flash (284B total parameters, 13B active), both using a Mixture-of-Experts architecture with novel Sparse Attention (DSA). The preview launched on April 24, 2026, with official GA releases of V4-Flash on July 31, 2026 and V4-Pro on August 12/13, 2026. Both models default to a 1-million-token context window, are openly available under the MIT license, and can be accessed via API and chat.deepseek.com.
Historique du momentum
21.05.19.08.
Fonctionnalités
| Key Benchmark (%) | SWE-bench Verified: 80.6% (V4-Pro-Max, vendor-reported); Artificial Analysis Intelligence Index: 53 (V4-Pro 0813, independently measured) |
| Context Window (Tokens) | 1,000,000 tokens (default across all official services), up to 384,000 tokens output |
| License | MIT License (open weights) |
| Multimodality | Text-based model with function calling and structured output; no native image/audio processing documented |
| Platform | API (OpenAI ChatCompletions & Anthropic-compatible), chat.deepseek.com (Web/App), Hugging Face Open Weights |
| Price per 1M Tokens | V4-Flash: $0.14 input / $0.28 output; V4-Pro: $0.435 input / $0.87 output (cache hits from $0.0028) |
| Release Date | Preview: April 24, 2026; V4-Flash official: July 31, 2026; V4-Pro GA: August 12/13, 2026 |