---
id: "entity-deepseek-v2"
type: "entity"
entityType: "product"
canonicalName: "DeepSeek v2"
aliases: ["DeepSeek-V2"]
source_timestamps: ["17:20:00"]
tags: ["product", "llm"]
related: ["concept-multi-head-latent-attention", "framework-memory-optimization-landscape"]
canonicalUrl: "https://platform.deepseek.com/docs"
sources: ["s49-killed-ram-limits"]
sourceVaultSlug: "s49-killed-ram-limits"
originDay: 49
---
# DeepSeek v2

DeepSeek v2 is an LLM noted in this source for introducing **Multi-Head Latent Attention (MLA)** — see [[concept-multi-head-latent-attention]].

**Architectural innovation**: MLA projects keys and values into a lower-dimensional latent space during training, structurally reducing the [[concept-kv-cache]] memory footprint **by design**. This makes it the canonical example of bucket #3 ('Architectural Redesign') in [[framework-memory-optimization-landscape]].

**Strategic contrast**: DeepSeek's approach (architectural, training-time) is complementary to Google's [[concept-turboquant]] (post-hoc, inference-time). They can stack.

**Canonical URL**: https://platform.deepseek.com/docs