Model

DeepSeek-V2.5

OPEN

DeepSeek · DeepSeek-V2 family · released Sep 5, 2024

Merged V2-Chat and Coder-V2 into a single 236B model, improving general writing/instruction-following while retaining coding strength.

ReasoningCodingVisionFunction callingTool useAgentic

1080.1

Elo · rank #194

Parameters

236B

Active params

21B (MoE)

Context

128K tokens

Architecture

MoE transformer, 236B total / 21B active — merge of DeepSeek-V2-Chat-0628 and DeepSeek-Coder-V2-Instruct-0724

License

DeepSeek Model License (code: MIT)

Languages

—

API price (in/out)

No hosted API

Modalities

text

Benchmark results

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

Aider PolyglotCoding17.8%#36

best: Claude Opus 4.5 · 89.4%

AlpacaEvalHuman preference50.5%#18

best: Tülu 2+DPO 70B · 95.1%

Arena-HardHuman preference76.2%#29

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

HumanEvalCoding89.0%#23

best: Claude Opus 4.5 · 99.4%

LiveCodeBenchCoding41.8%#131

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

MMLU-ProKnowledge65.8%#105

best: Claude Fable 5 · 91.5%

MT-BenchHuman preference9.02#9

best: Hunyuan-Large (A52B) · 9.4

SWE-bench VerifiedCoding16.8%#114

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-V2 family

Elo progression across releases

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

Resources

HFdeepseek-ai/DeepSeek-V2.5 DOCSapi-docs.deepseek.com/news/news0905 APIAvailable via DeepSeek API

API price weights · each benchmark row carries its own source badge (see methodology)