Model

DeepSeek-V2 236B

OPEN

DeepSeek · DeepSeek-V2 family · released May 7, 2024

Introduced Multi-head Latent Attention; its ultra-low launch price (~$0.14/$0.28 per M tokens) triggered China's 2024 LLM price war.

ReasoningCodingVisionFunction callingTool useAgentic

860.3

Elo · rank #245

Parameters

236B

Active params

21B (MoE)

Context

128K tokens

Architecture

MoE transformer, 236B total / 21B active — DeepSeekMoE + Multi-head Latent Attention (MLA)

License

DeepSeek Model License (code: MIT)

Languages

—

API price (in/out)

$0.14 / $0.28

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

AlpacaEvalHuman preference38.9%#28

best: Tülu 2+DPO 70B · 95.1%

BIG-Bench HardReasoning79.7%#33

best: ERNIE 4.5 300B-A47B · 94.3%

C-EvalKnowledge78.0%#35

best: Qwen3.6-Plus · 93.3%

CMMLUKnowledge81.6%#12

best: Doubao-1.5-Pro · 90.9%

GSM8KMath92.2%#29

best: Llama 3.1 405B · 96.8%

HumanEvalCoding81.1%#54

best: Claude Opus 4.5 · 99.4%

LiveCodeBenchCoding32.5%#145

best: DeepSeek-V4-Pro (Think Max) · 93.5%

MATH-500Math53.9%#121

best: GPT-5 · 99.4%

MBPPCoding72.0%#33

best: Llama-3.3-Nemotron-Super-49B v1 (Reasoning On) · 91.3%

MMLUKnowledge77.8%#108

best: OpenAI o3 · 92.9%

MT-BenchHuman preference8.97#12

best: Hunyuan-Large (A52B) · 9.4

Run it locally

VRAM @ Q4

118 GB

VRAM @ FP16

472 GB

Fits on (Q4)

M3 Ultra 512GBH200 141GBB200 192GB

236B total params exceed a single RTX 4090's 24GB even at 4-bit; requires multi-GPU or CPU/RAM offload

Quantizations

GGUF

DeepSeek-V2 family

Elo progression across releases

DeepSeek-V2 236BMay 2024860.3 DeepSeek-V2.5Sep 20241080.1

Resources

HFdeepseek-ai/DeepSeek-V2-Chat GHdeepseek-ai/DeepSeek-V2 DOCSapi-docs.deepseek.com APIAvailable via DeepSeek API

API price $0.14/$0.28 · each benchmark row carries its own source badge (see methodology)