length-controlled win rate on AlpacaEval 2.0 (Dubois et al., 2024), 8.97 overall score on MT-Bench (Zheng et al., 2023), and 7.91 overall score on AlignBench (Liu et al., 2023). The English open-ended conversation Testing DeepSeek-V2 Base 128K Context via "Needle In A HayStack" 1 2 3 4 5 6 7 8 9 10 Score Figure 4 | Evaluation results on the “Needle In A Haystack” (NIAH) tests. DeepSeek-V2 performs well 0.606 BBH (EM) 3-shot 68.7 59.9 78.9 81.0 78.9 MMLU (Acc.) 5-shot 71.3 77.2 77.6 78.9 78.5 DROP (F1) 3-shot 69.7 71.5 80.4 82.5 80.1 ARC-Easy (Acc.) 25-shot 95.3 97.1 97.3 97.9 97.6 ARC-Challenge (Acc

‘The AI Index 2025 Annual Report,’ AI Index Steering Committee, Stanford HAI (4/25) LMSYS Arena Score AI Model Compute Costs High / Rising + Inference Costs Per Token Falling = Performance Converging its first attempt. Source: Epoch AI (5/25) DeepSeek R1 (1/25) scored 93% vs. o3- mini’s (1/25) score of 95% Non-Downloadable (Closed) Downloadable (Open) AI Monetization Threats = Rising Competition Threats = Rising Competition + Open-Source Momentum + China’s Rise Artificial Analysis Quality Index Score 0 50 100 Coding Quantitative Reasoning Reasing & Knowledge Scientific Reasoning & Knowledge

Frequency & High Compute Efficiency ˃ Supported on U200 – 3 Instances U250 – 4 Instances Amazon F1 ˃ ~1536 DSPs @ 700MHz Execution Controller Spill / Restore DMA Controller Weights DMA Controller

Understudy for Gisting Evaluation). 3. Select the instruction candidate with the highest evaluation score. This candidate will be the final prompt you can use in your software application or chatbot. You