DeepSeek Coder V2 Lite Instruct

DeepSeek Coder V2 Lite Instruct is a 15.71-billion-parameter Mixture of Experts model from DeepSeek, optimized for code generation and mathematical reasoning. It activates 6 of 64 experts plus 2 shared experts per token, delivering strong coding performance with low compute cost. It supports 338 programming languages and nine natural languages, making it one of the broadest polyglot code models available. With a 160K context window and flash attention, it handles large codebases efficiently and quantizes well to GGUF for self-hosted deployment.

Mixture-of-Experts

This is a Mixture-of-Experts model. MoE models load all parameters but activate only a fraction per token — learn why this changes VRAM planning, quantization, and deployment strategy.

Read the MoE deployment guide →

Hardware Configuration

Vendor Select vendor...

Product —

Platform Select platform...

Family —

Model —

VRAM —

System RAM (GB) Optional — for precise deployment recommendations

Clear Configuration

Quantization	Quality	Size	Fit
Q8_0	High	15.56 GB	—
Q6_K	High	13.1 GB	—
Q5_K_M	Medium	11.04 GB	—
Q4_K_M	Medium	9.65 GB	—
Q3_K_L	Low	7.88 GB	—

Last updated: March 5, 2026