DeepSeek-Coder-V2-Instruct
DeepSeekDedicated coding mixture-of-experts model, instruction-tuned for generation, completion and chat, built on a DeepSeek-V2 checkpoint with six trillion additional code tokens. It ships in a 236B-total variant with 21B active parameters and a 16B Lite variant, supports 338 programming languages and a 128K context.
A 2024-generation coder, superseded by the V3 and V4 lines. The code repository and the weights carry different terms; the openness axis follows the ones on the weights. Verified 2026-08-13 via the model card and the model-license body.
Openness
3 high confidence- weights
- open(downloadable on HF)
- data
- closed
- code
- open(MIT, repo code)
- model-license
- DeepSeek-Model-License(non-OSI
MIT-licensed inference code and downloadable weights under the DeepSeek Model License, with no training corpus released, so this is open weights. The license's Attachment A forbids military, illegal and discriminatory use and requires those terms to propagate to derivatives, but permits commercial use at any scale - a restriction on conduct rather than on commerce, which is why it does not pull the score down to restricted. DeepSeek's pretrained base models are treated the same way on the same license, and the result is consistent with the DeepSeek chat tier, which reaches the same score under plain MIT.
- https://huggingface.co/deepseek-ai/DeepSeek-Coder-V2-Instruct recorded 2026-08-13
License section - "This code repository is licensed under the MIT License. The use of DeepSeek-Coder-V2 Base/Instruct models is subject to the Model License"; `"gated":false` in the embedded repo state, so the 236B-A21B MoE weights download without a barrier; no training corpus published on the card
- https://github.com/deepseek-ai/DeepSeek-Coder-V2/blob/main/LICENSE-MODEL recorded 2026-08-13
DEEPSEEK LICENSE AGREEMENT Version 1.0, 23 October 2023 - preamble states it "added use-based restrictions"; Attachment A carries them; non-OSI custom license
Adoption
3 high confidenceThe family bands on the sum of downloads across its SKUs. The flagship 236B SKU reads 7,143 in the trailing 30 days and the Lite SKU declared alongside it reads 587,824, summing to 594,967, which lands in 100K-1M. The line has been superseded by DeepSeek-V2.5/V3/R1 and by the 2026 coders and the flagship is in real decline, so the band rests on the Lite SKU people actually run.
- https://huggingface.co/api/models/deepseek-ai/DeepSeek-Coder-V2-Instruct recorded 2026-08-14
downloads field = 7,143 in the trailing 30 days; 698 likes; gated false
- https://huggingface.co/api/models/deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct recorded 2026-08-14
downloads field = 587,824 in the trailing 30 days; 631 likes; gated false
Capability
3 medium confidenceGPT-4-Turbo-class open coder at its 2024 release, but a 2024-generation model now mid-tier vs 2026 frontier coders (Qwen3-Coder, DeepSeek-V4, GLM-5.1, Kimi K2.6). HF card defers exact HumanEval/MBPP/SWE-bench numbers to the paper; scored 3 with medium confidence on vendor positioning.
- https://huggingface.co/deepseek-ai/DeepSeek-Coder-V2-Instruct recorded 2026-08-13
vendor claim of performance superior to GPT4-Turbo, Claude 3 Opus and Gemini 1.5 Pro on coding and math; support expanded from 86 to 338 programming languages, context 16K -> 128K
Verified 2026-08-13