LANGUAGE MODEL DeepSeek Last updated:

DeepSeek V3

671B Open MoE at $5.5M Training Cost

DeepSeek's December 2024 frontier release: a 671-billion-parameter Mixture-of-Experts model approaching GPT-4o quality, with weights open and a published training cost of $5.5M — one to two orders of magnitude below US frontier estimates. Quietly released on December 26, 2024, it would become the foundation for R1 four weeks later and the source of the January 2025 Nvidia stock shock.

Intelligence
Below avg
Cost
Low
$0.40 in / $0.89 out
Context
131K
Up to 131,072 tokens
How are Intelligence, Speed & Cost bucketed?
Intelligence and Speed buckets are percentile ranks on Artificial Analysis. Cost buckets are fixed dollar thresholds keyed off output-token price ($/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 V3, not R1, is the model that actually changed the economics. R1 added the reasoning headlines, but V3 was where the cost shock originated. Every subsequent debate about US export controls, AI capex sustainability, and Chinese AI capability has V3 as a central data point.

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.
Agent Workflows
Built for tool use and autonomous tasks.

Context Window

131k tokens
≈ 101 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
Available
Product / App
Available
Open Source
Released
Enterprise
Contact sales

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
2.4
Coding
SciCode · scaled to 10
1.8
4.3
3.9
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
6.1
Cost efficiency
Input price ($/M tokens) · cheaper scores higher
6.2
10.0
10.0
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

  • 671B-param MoE (37B active per token) — the open-weights model that finally matched GPT-4 / Claude 3.5 Sonnet quality
  • 82.6% HumanEval (coding) and 90.2% MATH-500 — outperformed GPT-4o, Claude 3.5 Sonnet, and Llama 3 on each
  • Trained for $5.576M on 2.788M H800 GPU hours — 1/100th of comparable proprietary training budgets
  • Auxiliary-loss-free load balancing was a real architectural innovation, not just scaling
  • MoE inference needs 8+ H100s; not laptop-friendly even though the weights are open
  • MMLU score of 88.5 was the highest open-source result at release

Best use cases

  • Self-hosted frontier-quality LLM for orgs that can't use proprietary APIs
  • Math, coding, and general knowledge benchmarks at open-source price
  • Fine-tuning and distillation into smaller production models
  • Cost-sensitive bulk inference via budget providers

Tools to try

Not ideal for

  • Edge / single-GPU deployments — the MoE size is large despite low active params
  • Multimodal use cases — text-only
  • Latency-sensitive interactive chat (better with smaller open models like Qwen 2.5)

Model Evolution

deepseek-v is DeepSeek's language model family.

View full evolution tree →