LLM ReferenceLLM Reference

DeepSeek V3

Open Source

About

DeepSeek V3: Latest flagship model. 685B total with MoE. 128K context. Open-source.

Capabilities

VisionMultimodalReasoningFunction CallingTool UseStructured OutputsCode Execution

Providers(11)

Compare all →
ProviderInput (per 1M)Output (per 1M)Type
DeepInfra$32$89Serverless
Fireworks AI$0.56$1.68Serverless
DeepSeek Platform$0.14$0.28Serverless
Microsoft FoundryServerlessProvisioned
OpenRouter$0.26$0.38Serverless
NVIDIA NIMServerless
AWS BedrockServerless
Together AIServerless
Bitdeer AI$0.1$0.3Serverless
SiliconFlow$0.15$0.5Serverless
Replicate API$1.45$1.45Serverless

Benchmark Scores(8)

BenchmarkScoreVersionSource
HellaSwag95.710-shotOpen LLM Leaderboard, DeepSeek official
HumanEval85.5pass@1Open LLM Leaderboard, DeepSeek official
Massive Multitask Language Understanding88.55-shotOpen LLM Leaderboard
LiveCodeBench49.62026-04https://livecodebench.github.io/performances_generation.json
Aider Polyglot48.42026-04https://aider.chat/docs/leaderboards
BigCodeBench50.02025-01 (Instruct Pass@1)https://bigcode-bench.github.io/results.json
Chatbot Arena1302.0https://lmarena.ai
MMLU PRO75.9https://huggingface.co/spaces/TIGER-Lab/MMLU-Pro

Rankings

Specifications

Released2024-12-26
Parameters671B
Context64k
ArchitectureMixture of Experts
Knowledge cutoff2024-04
Specializationgeneral
Trainingfinetuning

Created by

Advancing artificial general intelligence (AGI).

Hangzhou, Zhejiang, China
Founded 2023
Website