LANGUAGE MODEL DeepSeek Last updated:

DeepSeek-V2

Open MoE with Multi-Head Latent Attention

A Chinese AI lab's 236B-parameter Mixture-of-Experts model, released open-weight with a new attention architecture (MLA) that dramatically reduced inference memory. Its API was priced at $0.28 per million output tokens — two orders of magnitude cheaper than GPT-4 Turbo. The pricing itself was the news: DeepSeek had broken the perceived floor of frontier-quality LLM inference pricing.

Intelligence
Below avg
Cost
Free
Open weights — self-host
Context
128K
Up to 128,000 tokens
How are Intelligence, Speed & Cost bucketed?
Intelligence and Speed buckets are percentile ranks on Artificial Analysis. Cost buckets are fixed dollar thresholds keyed off output-token price ($/M out).
Intelligence
  • Top 1%≤ 1%
  • Top 5%≤ 5%
  • Top 10%≤ 10%
  • Good≤ 25%
  • Medium≤ 50%
  • Below avg> 50%
Speed
  • Top 1%≥ 345 tok/s
  • Top 5%≥ 237 tok/s
  • Top 10%≥ 196 tok/s
  • Good≥ 146 tok/s
  • Medium≥ 90 tok/s
  • Slow< 90 tok/s
Cost
  • Freeopen weights · self-host
  • Low< $1 / M out
  • Moderate$1–5 / M out
  • High≥ $5 / M out

Why it matters

DeepSeek V2 is the proximal cause of the global LLM inference price collapse in 2024-25. Every frontier lab now has to explain why they charge what they do — when a Chinese open model runs at $0.28/1M tokens and meets 90% of enterprise quality needs.

Core Capabilities

Long Documents
Handles entire codebases, books, and multi-doc RAG.
Generative
Produces images, video, audio, or other media.
Research
Foundational paper or scientific contribution.

Context Window

128k tokens
≈ 98 pages
4k Chat 聊天
32k Long docs 长文档
128k This model 本模型
400k Multi-doc 多文档
1M Codebase 整个代码库
10M

Availability

API
Not available
Product / App
Not available
Open Source
Released
Enterprise

Pricing Model

Free / self-host
Open weights — pay only for compute.
Self-host

Capability / Performance

Where this model sits relative to the middle 60% of models in the tree. All scores are 0–10 (higher is better).

Lower 20% Upper 80% This model
Reasoning
AA Intelligence Index · scaled to 10
1.7
5.6
1.8
Coding
SciCode · scaled to 10
1.8
4.3
4.4
Agentic tasks
Terminal-Bench Hard · scaled to 10
0.2
3.6
3.6
Context / memory
Context window size · log-scaled
6.0
9.0
6.0
Lower 20% 20th percentile — 20% of models score below this This model Where the current model lands Upper 80% 80th percentile — only 20% of models score above this Percentile boundaries are computed across every model in the tree that reports the underlying benchmark for each capability.

What it feels like

  • First DeepSeek model that the wider OSS community took seriously — 236B MoE with only 21B active per token
  • Saved 42.5% training cost vs DeepSeek 67B, reduced KV cache 93.3%, boosted max throughput 5.76x
  • DeepSeekMoE architecture (fine-grained experts + shared expert isolation) became the template
  • 128K context window at open-weights pricing — rare in mid-2024
  • MMLU performance in the top open-source tier despite the small active-parameter footprint
  • Foreshadowed V3 / R1 — proved the cost-efficient training playbook before it shocked Wall Street

Best use cases

  • Self-hosted Chinese / multilingual deployments where API access was politically risky
  • Cost-sensitive bulk inference via budget providers
  • MoE-architecture research and KV-cache optimisation experiments
  • Long-context retrieval workflows (128K) at open-weights price

Tools to try

Not ideal for

  • Frontier reasoning by 2025 — superseded by V3 / V3.1 / V4 generations
  • Edge / single-GPU deployments (236B MoE still demands multi-GPU serving)
  • Vision / multimodal tasks (DeepSeek-VL came separately)

Model Evolution

View full evolution tree →