Repo-Level Coding
π Reading Time: ~15 min read
π
Updated 2026-03
DeepSeek-Coder-V2: 338 Languages Support, 128K Context & Repository-Level Inference
MoE coding powerhouse DeepSeek-Coder-V2 (16B/236B). Benchmark analysis, FIM middle code infilling, and multi-file code completion via Tabby and Continue.
π» MoE transition of DeepSeek-Coder-V2
DeepSeek-Coder-V2 is built on the DeepSeek-V2 base architecture and includes two specifications: 16B (activation 2.4B) and 236B (activation 21B). It supports up to 338 programming languages ββand configuration file formats, and its mathematical reasoning and long code completion benchmark scores are on par with GPT-4o.
β‘ High-speed image pulling and vLLM deployment
# Image pull Lite 16B version (suitable for a single 16G/24G graphics card)
export HF_ENDPOINT="https://hf-mirror.net"
huggingface-cli download deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct \
--local-dir /data/models/DeepSeek-Coder-V2-Lite-Instruct \
--local-dir-use-symlinks False
# Start vLLM instance
vllm serve /data/models/DeepSeek-Coder-V2-Lite-Instruct \
--max-model-len 32768 \
--trust-remote-code \
--port 8000