DeepSeek-Coder
The original DeepSeek code series — 1.3B to 33B trained on 2T tokens, 87% code.
About
DeepSeek-Coder (November 2023) is the original DeepSeek code model series — 1.3B, 5.7B, 6.7B, and 33B, each in base and instruct versions — trained from scratch on 2 trillion tokens (87% code, 13% natural language) with a 16K project-level window and fill-in-the-middle support. It was the first open coder to clearly beat CodeLlama across benchmarks and put DeepSeek on the map before V2, V3, and R1. Weights carry the custom DeepSeek license (non-OSI; commercial use allowed within its terms). Superseded by DeepSeek-Coder-V2 and the reasoning line, but the 6.7B remains a handy self-hosted coder.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Team-reported: state-of-the-art among publicly available code models at release across HumanEval, MultiPL-E, MBPP, and DS-1000.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (verified tier, metrics-v1). Evidence: DeepSeek-Coder paper (arXiv 2401.14196) + HF README read this session: Nov 2023, 1.3B/5.7B/6.7B/33B, 2T tokens (87% code), 16K window, FIM.; Custom DeepSeek license (non-OSI, commercial allowed within terms) explicit. Distinct from the catalog's deepseek-coder-v2 entry..