Activefrontierllm Open Source

DeepSeek Coder V2

by DeepSeek· Released June 2024· Cutoff February 2024

DeepSeek Coder V2 is an open-source Mixture-of-Experts (MoE) code language model that achieves performance comparable to GPT-4 Turbo in coding tasks. It supports 338 programming languages and features a 128K context window, making it one of the most capable code-specific models available.

Official Site API Docs 🤗 Hugging Face 📄 Research Paper

Input cost

$0.14 per 1M tokens

Output cost

$0.28 per 1M tokens

Context window

128K tokens

Max output

—

Modalities

text

Parameters

236B (total), 21B (active)

License

MIT

Capabilities

Code GenerationCode CompletionCode InfillingMulti-language SupportFunction CallingStreamingJSON Mode

Best For

Advanced code generation, completion, and infilling across a wide range of programming languages.

Strengths

Top-tier coding performance comparable to GPT-4 Turbo
Supports 338 programming languages
128K context window for large codebases
Open-source with permissive license
Cost-effective pricing

Limitations

Primarily focused on code, less optimized for general text tasks
MoE architecture may have higher memory footprint
Limited non-English language support
No vision or multimodal capabilities

Use Cases

Automated code generation and completion

Code review and bug fixing

Documentation generation from code

Code translation between programming languages

Educational tool for learning programming

Refactoring and code optimization

Building developer tools and assistants

Improvements Over Previous Model

Context window increased from 16K to 128K tokens
Parameter count increased from 33B to 236B (total) / 21B (active)
Supports 338 programming languages (up from 86)
Coding benchmark scores improved significantly (HumanEval 90.2% vs 73.8%)
Added support for function calling and JSON mode
Open-source under MIT license (previous version was Apache-2.0)

Back to all models