Model

DeepSeek LLM 67B

OPEN

DeepSeek · DeepSeek LLM family · released Nov 29, 2023

First general-purpose DeepSeek foundation model (dense, 7B/67B); 67B surpassed Llama 2 70B on code, math and reasoning.

ReasoningCodingVisionFunction callingTool useAgentic

482.1

Elo · rank #310

Parameters

67B

Active params

67B (dense)

Context

4K tokens

Architecture

Dense auto-regressive transformer decoder with grouped-query attention, 4K context

License

DeepSeek Model License (custom, commercial use permitted)

Languages

API price (in/out)

No hosted API

Modalities

text

Benchmark results

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

AGIEvalReasoning46.4%#23

best: OLMo 3-Think 32B · 88.2%

ARC-ChallengeReasoning64.1%#71

best: Llama 3.1 405B · 96.9%

ARC-EasyReasoning81.6%#22

best: Phi-3-medium (14B) · 97.7%

BIG-Bench HardReasoning71.7%#48

best: ERNIE 4.5 300B-A47B · 94.3%

C-EvalKnowledge65.2%#43

best: Qwen3.6-Plus · 93.3%

CMMLUKnowledge67.8%#20

best: Doubao-1.5-Pro · 90.9%

DROPReasoning71.9#42

best: Hunyuan-T1 · 93.1

GPQA DiamondReasoning29.8%#264

best: GPT-5.6 · 94.6%

GSM8KMath84.1%#70

best: Llama 3.1 405B · 96.8%

HellaSwagReasoning75.7%#110

best: Claude 3 Opus · 95.4%

HumanEvalCoding73.8%#70

best: Claude Opus 4.5 · 99.4%

IFEvalReasoning50.8%#133

best: Gemma 4 26B A4B · 98.5%

MATH-500Math32.6%#157

best: GPT-5 · 99.4%

MBPPCoding61.4%#56

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

MMLU-ProKnowledge39.4%#153

best: Claude Fable 5 · 91.5%

MMLUKnowledge71.1%#139

best: OpenAI o3 · 92.9%

MT-BenchHuman preference8.35#34

best: Hunyuan-Large (A52B) · 9.4

OpenBookQAReasoning63.2%#11

best: Claude 1 · 90.8%

PIQAReasoning82.6%#24

best: GPT-4o mini · 93.1%

RACE-HReasoning56.0%#10

best: Claude 3 Opus · 92.9%

TriviaQAKnowledge81.5%#12

best: Sarvam-1 (2B) · 90.6%

WinoGrandeReasoning76.0%#58

best: PaLM 2 · 90.9%

Run it locally

VRAM @ Q4

39 GB

VRAM @ FP16

135 GB

Fits on (Q4)

M3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB

67B exceeds a single RTX 4090's 24GB VRAM even at Q4 (~39GB); no verified single-4090 llama.cpp benchmark found.

Quantizations

GGUF · AWQ · GPTQ

DeepSeek LLM family

Elo progression across releases

DeepSeek LLM 67BNov 2023482.1

Resources

HFdeepseek-ai/deepseek-llm-67b-chat GHdeepseek-ai/DeepSeek-LLM APIAvailable via DeepSeek API

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