Model

DeepSeek-Coder-V2 236B

OPEN

DeepSeek · DeepSeek-Coder-V2 family · released Jun 17, 2024

Reached GPT-4-Turbo-level coding (90.2 HumanEval) while expanding language support from 86 to 338 programming languages at 128K context.

ReasoningCodingVisionFunction callingTool useAgentic

1012.1

Elo · rank #210

Parameters

236B

Active params

21B (MoE)

Context

128K tokens

Architecture

MoE transformer, 236B total / 21B active — DeepSeekMoE + MLA, further pretrained from a DeepSeek-V2 checkpoint on 6T additional code/math tokens

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

AIMEMath13.3%#146

best: GPT-5.2 · 100.0%

AlpacaEvalHuman preference36.9%#29

best: Tülu 2+DPO 70B · 95.1%

ARC-ChallengeReasoning92.8%#18

best: Llama 3.1 405B · 96.9%

Arena-HardHuman preference65.0%#37

best: Qwen3-235B-A22B (Non-Thinking) · 96.1%

BIG-Bench HardReasoning83.9%#21

best: ERNIE 4.5 300B-A47B · 94.3%

C-EvalKnowledge79.4%#34

best: Qwen3.6-Plus · 93.3%

CMMLUKnowledge80.9%#13

best: Doubao-1.5-Pro · 90.9%

GSM8KMath94.9%#10

best: Llama 3.1 405B · 96.8%

HumanEvalCoding90.2%#17

best: Claude Opus 4.5 · 99.4%

LiveCodeBenchCoding43.4%#127

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

MATH-500Math75.7%#78

best: GPT-5 · 99.4%

MBPP+Coding76.2%#9

best: Llama 3.1 405B · 88.6%

MMLU-ProKnowledge63.6%#110

best: Claude Fable 5 · 91.5%

MMLUKnowledge79.2%#96

best: OpenAI o3 · 92.9%

MT-BenchHuman preference8.77#17

best: Hunyuan-Large (A52B) · 9.4

SWE-bench VerifiedCoding12.7%#116

best: Claude Fable 5 · 95.0%

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-Coder-V2 family

Elo progression across releases

DeepSeek-Coder-V2 236BJun 20241012.1

Resources

HFdeepseek-ai/DeepSeek-Coder-V2-Instruct GHdeepseek-ai/DeepSeek-Coder-V2 APIAvailable via DeepSeek API

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