DeepSeek-V3

V3 (Dec 2024 Baseline)

Checked on 15 Nov 2026 against DeepSeek API docs.

The baseline foundation model that established DeepSeek's global engineering leadership. Trained from scratch on 14.8T tokens, introducing Multi-Head Latent Attention (MLA), DeepSeekMoE with auxiliary-loss-free load balancing, and native FP8 mixed precision.

API list price
Not listed on the current DeepSeek API pricing page (deepseek-flash and deepseek-v4-pro only).
Context
128K Tokens
Parameters
671B Total; active 37B Active / Token
Architecture
DeepSeekMoE (256 routed + 1 shared expert) + MLA + FP8 Mixed Precision

Not listed on the current DeepSeek API pricing page (deepseek-flash and deepseek-v4-pro only).

Sourced benchmarks

  • MMLU (Massive Multitask): 88.5% — Standard benchmark testing factual and professional competency. Source
  • HumanEval-Mul (Coding): 82.6% — HumanEval-Mul pass@1 on the DeepSeek-V3 instruct evaluation table. Source
  • MATH-500: 90.2% — Multi-step mathematical problem solving. Source
  • GSM8K (Grade School Math): 89.3% — Grade-school math word problems. Source

DeepSeek-V3 in the API timeline

DeepSeek-V3 is the December 2024 foundation MoE described in the DeepSeek-V3 README. It introduced Multi-Head Latent Attention (MLA), DeepSeekMoE with auxiliary-loss-free load balancing, and native FP8 mixed-precision training on 14.8T tokens. The README's evaluation table lists MMLU 88.5, HumanEval-Mul pass@1 82.6, and GSM8K 89.3 for the instruct checkpoint.

On the consumer API, DeepSeek later routed the legacy alias deepseek-chat to newer backends (see the 2024-12-26 change log mirror and the model name guide). The current Models & Pricing table centers on deepseek-flash and deepseek-v4-pro; V3 remains relevant for weight downloads, architecture study, and historical comparisons.

Training efficiency claims in the README — for example MLA KV-cache compression and MoE routing — underpin the open-source kernels catalogued on /open-source (FlashMLA, DeepEP, DeepGEMM). This directory does not publish a separate V3 API price row today; budget using the live Flash and Pro rows on /api-pricing or the cross-vendor comparison.

All models · API pricing · Off-peak pricing guide · LLM API price comparison · Model names · FAQ