DeepSeek-Coder-V2

// Overview

DeepSeek-Coder-V2 is an open-weight, code-specialized Mixture-of-Experts model from DeepSeek, released on June 17, 2024. It comes in two sizes: a flagship with 236B total parameters and 21B active, and a smaller DeepSeek-Coder-V2-Lite with 16B total and 2.4B active. Both ship as base and instruction-tuned checkpoints, and were further pre-trained from an intermediate DeepSeek-V2 checkpoint on an additional 6 trillion tokens of code and math data.

Compared with the original DeepSeek Coder, DeepSeek-Coder-V2 expands programming-language coverage from 86 to 338 languages and extends the context window from 16K to 128K tokens. DeepSeek positions it as the first open-source model to break the barrier of closed-source code intelligence: on coding and math benchmarks it reaches performance comparable to GPT-4-Turbo and reportedly edges out Claude 3 Opus and Gemini 1.5 Pro of that era.

On reported benchmarks the 236B Instruct model scores 90.2% on HumanEval, 76.2% on MBPP+, 43.4% on LiveCodeBench, 75.7% on MATH, and 94.9% on GSM8K, while keeping solid general-language ability at 79.2% MMLU. The weights are downloadable from Hugging Face under an MIT code license plus a DeepSeek Model License that permits commercial use; the hosted API endpoint was later merged into DeepSeek-V2.5 in September 2024.

// Benchmarks

1. HumanEval (DeepSeek-Coder-V2-Instruct)90.2%
2. MBPP+ (EvalPlus)76.2%
3. LiveCodeBench43.4%
4. SWE-Bench12.7%
5. Aider (code editing)73.7%
6. MATH75.7%
7. GSM8K94.9%
8. MMLU79.2%
9. Arena-Hard65%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Weights are downloadable from Hugging Face for self-hosting; the original hosted deepseek-coder API endpoint was merged into DeepSeek-V2.5 in September 2024.

Pricing source ↗

// Strengths

• GPT-4-Turbo-class coding performance from a fully open-weight model (90.2% HumanEval)
• Broad language coverage — 338 programming languages, up from 86 in DeepSeek Coder v1
• 128K-token context for whole-repository and long-file reasoning
• Strong mathematical reasoning: 75.7% MATH and 94.9% GSM8K
• Two sizes — a 236B/21B-active flagship and a lightweight 16B/2.4B-active Lite — plus base and instruct checkpoints
• Permissive licensing (MIT code + commercial-use model license) for self-hosting

// Best for

• Code generation and autocompletion across many languages
• Code repair and bug fixing
• Mathematical and algorithmic reasoning
• Repository-scale code understanding using the 128K context
• Self-hosted, on-prem coding assistants where open weights are required
• Cost-efficient open-weight alternative to closed coding APIs

// How to access

// DeepSeek Coder — every version

The full lineage of the DeepSeek Coder line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.