Atlas

Models

← All models

DeepSeek-V2-Chat

DeepSeek · 2024-05-06 · 235.7B parameters

DeepSeek-V2-Chat is the reinforcement-learning-aligned chat checkpoint of DeepSeek-V2, released May 6, 2024; its downloadable weights contain exactly 235,741,434,880 parameters, while DeepSeek reports only rounded figures of 236B total and 21B activated per token. The MLA/DeepSeekMoE base was pretrained on a reported 8.1T tokens, extended with YaRN from 4K to documented 128K context support, then aligned on 1.5M SFT instances and two-stage GRPO; its configuration declares 163,840 maximum positions.

Benchmark scores

BenchmarkScore
Artificial Analysis Intelligence Index3.6
BigCodeBench Complete49.0
BigCodeBench Instruct40.4
BigCodeBench-Hard Complete15.5
BigCodeBench-Hard Instruct21.6
Chinese996
Loading Atlas data…