sameerhfedu/deepseek-v4-flash

DeepSeek-V4-Flash Interactive Demo & Reasoning Playground

Experience DeepSeek-V4-Flash (284B total / 13B activated parameters) with 1 Million token context, Hybrid Attention (CSA + HCA), and multi-mode reasoning (Non-Think, Think High, and Think Max).

Features

• —⚡ Real-Time Streaming Chat: Instant token streaming powered by the Hugging Face Router API.
• —🧠 Multi-Mode Reasoning Engine: Switch seamlessly between Non-Think, Think High, and Think Max deliberation modes with live <think> inspection.
• —📐 LaTeX & Math Support: Real-time formula rendering with KaTeX.
• —💻 Syntax-Highlighted Code: Beautiful code blocks with one-click copy and language detection.
• —💡 Interactive Preset Library: Instant benchmark riddles, coding challenges, math proofs, and architecture demos.
• —🔬 Architecture & Benchmark Deep Dive: Comprehensive breakdown of MoE routing, mHC, Muon optimization, and evaluation metrics.

Model Details

• —Model ID: `deepseek-ai/DeepSeek-V4-Flash`
• —Architecture: MoE (284B total / 13B active per token, 6 experts/token)
• —Context Length: 1,000,000 tokens (1M)
• —Precision: FP4 (experts) + FP8 (dense layers) Mixed
• —Paper: arXiv:2606.19348