DeepSeek V5 (Leaked Staging)
Autonomous coding agents scoring 81.4% on SWE-bench Verified, zero-telemetry self-hosting, sub-cent 1M-token processing, and formal mathematical proofs.
SWE-bench Verified
81.4%
MMLU-Pro
87.5%
MATH-500
96.8%
Context Window
1,000,000 tokens
| Specification | Value |
|---|---|
| Developer / Lab | DeepSeek AI |
| Parameters Total | 1.8 Trillion (Sparse MoE) |
| Parameters Active | 128 Billion Active Parameters |
| Architecture Type | Multi-Head Latent Attention 2.0 (MLA 2.0) + 128-Expert MoE (8 Active) + DualPipe 2.0 + Native Hybrid Reasoning |
| License | Open-Weights MIT License (Hugging Face / ModelScope / DeepSeek API) |
| Input Pricing (per 1M) | $0.12 |
| Output Pricing (per 1M) | $0.48 |
Head-to-Head Comparisons with DeepSeek V5 (Leaked Staging)
DeepSeek V5 (Leaked Staging) vs Gemini 4 Argon →DeepSeek V5 (Leaked Staging) vs GPT-6.1 Sol →DeepSeek V5 (Leaked Staging) vs Gemini 4 Pro (Leaked Benchmark) →DeepSeek V5 (Leaked Staging) vs Space Bunny Alpha (Stealth) →DeepSeek V5 (Leaked Staging) vs GPT-6 Sol →DeepSeek V5 (Leaked Staging) vs GPT-6 Luna →