DeepSeek Flash
#43 en Modèles de langage open sourceDeepSeek · flash · depuis 31. Juli 2026 (offizielle GA-Version "V4-Flash-0731"; ursprüngliche Preview: 24. April 2026) · 3× · vu le 15 août 2026
DeepSeek Flash (officially DeepSeek-V4-Flash, "0731" build) is an open-source language model from DeepSeek positioned as the cost-efficient variant of the DeepSeek-V4 model family. It uses a Mixture-of-Experts architecture with 284 billion total parameters, of which 13 billion are activated per token, and supports a 1-million-token context window. The official version released on July 31, 2026 substantially improves agentic and coding capabilities over the April preview through extensive post-training, without changing the underlying architecture. Weights are freely available on Hugging Face under the MIT License.
Fonctionnalités
| Key Benchmark (%) | Terminal-Bench 2.1: 82.7 (official jump from 61.8 to 82.7 via post-training) |
| Context Window (Tokens) | 1,000,000 tokens (up to 384,000 max output tokens) |
| License | MIT License (open weights) |
| Multimodality | No evidence of multimodality found; model is primarily text-based with function calling/structured output |
| Platform | DeepSeek API (deepseek-v4-flash), chat.deepseek.com, Hugging Face, LM Studio, Ollama |
| Price per 1M Tokens | $0.14 input (cache miss) / $0.0028 input (cache hit) / $0.28 output |
| Release Date | July 31, 2026 (official 0731 release); preview on April 24, 2026 |