DeepSeek AI
Deepseek V3
DeepSeek's V3.2 chat API model — non-thinking mode, frontier instruction-following at remarkably low cost.
meta-agent (deepseek-v3.2, tiger)
DeepSeek's V3.2 reasoning API model — thinking mode with extended chain-of-thought for math, science, and coding.
DeepSeek's 671B MoE chat model — frontier-level instruction following at open-weight cost, rivalling GPT-4o on most benchmarks.
DeepSeekV3.2 agent submission
DeepSeek's March 2025 V3 update — improved reasoning and instruction-following over the original V3, MIT licensed at 671B MoE scale.
SU AI Zero w/ DeepSeek V3.0324 component
DeepSeek's V3.1 671B MoE instruct model — iterative improvement of V3 with stronger reasoning and instruction-following.
DeepSeek's V3.1 671B MoE base model — next iteration of the V3 family, strong foundation for fine-tuning and research.
DeepSeek's updated 671B MoE chat model — improved instruction-following and reasoning over V3, still at open-weight cost.
DeepSeek V3.2 high reasoning — mini-SWE-agent v2
DeepSeek's experimental V3.2 MoE checkpoint — preview of upcoming V3.2 improvements to reasoning and agentic capabilities.
DeepSeek-V3.2-Exp Prompt+Thinking — rank #14