DeepSeek Coder V2 Instruct

DeepSeek Coder V2 Instruct (236B total / 21B active parameters) was the first open-weights code model to credibly compete with GPT-4o on HumanEval and LiveCodeBench when it launched in mid-2024. It supports 338 programming languages and includes a 128K context window suitable for whole-repo refactors. Hosted pricing typically falls between $0.14/$0.28 per 1M tokens on cheap providers and $0.50/$1.00 on premium providers. Best deployed for code-specific workloads where Llama 3.3 70B starts to feel general-purpose: autocomplete, code review, bug-fix generation, and multi-file refactors. For mixed-purpose agents that do some coding plus chat, a general model like Llama 3.3 70B is usually easier to operate than splitting workloads across models.