deepseek-ai/DeepSeek-V3

The inference module now supports the DeepSeek-V3 architecture, introducing configuration files for 16B, 236B, and 671B parameter models, alongside a new \generate.py\ script for interactive and batch text generation. The implementation includes a \convert.py\ script to transform HuggingFace checkpoints into the required format and a \fp8\_cast\_bf16.py\ utility to convert FP8 weights to BF16. The core \model.py\ and \kernel.py\ files implement the transformer logic, including Multi-Head Latent Attention (MLA) and Mixture of Experts (MoE) routing, with support for FP8 quantization via Triton-based GEMM kernels.

inference · high confidence

The repository now includes the DeepSeek-V3 model, a 671B parameter Mixture-of-Experts (MoE) language model. The update introduces a new README.md featuring the model's architecture, evaluation results, and download links, alongside a dedicated README\_WEIGHTS.md detailing the 671B main model and 11.5B Multi-Token Prediction (MTP) modules. Additionally, the repository now provides the MIT Code License (LICENSE-CODE) and the DeepSeek Model License (LICENSE-MODEL), which includes specific use-based restrictions. A .gitignore file is also added to manage build artifacts and environment files.

(repo-wide) · high confidence

Written by watchdog.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

• CAI 46 → 48 (+1.7)
• Rubric changed (rubric-2026.08.15 → rubric-2026.09.15) — scores are not directly comparable.

• Code Health 97 → 97 (-0.3)
• Architecture 69 → 69 (+0.0)
• Maturity 49 → 53 (+4.5)
• Readiness 25 → 26 (+0.4)
• Security 92 → 96 (+4.2)

• Dimension evaluation failed
• No exposed public API
• The table of contents lists 'Model Summary' but no actual summary of what the model does or how it differs from prior models. (README.md)

• Critical CVE: [GHSA redacted] (inference/requirements.txt)
• Documentation: no installation or build instructions (README.md)
• Documentation: no project overview (README.md)
• Documentation: no usage examples (README.md)
• High CVE: [GHSA redacted] (inference/requirements.txt)
• No dependency advisory monitoring
• Orphaned files with no living knowledge
• Outdated: safetensors
• Outdated: torch
• Outdated: transformers
• Outdated: triton
• Workflow token permissions not restricted
• convert.main (cognitive 25) (inference/convert.py)
• fp8_cast_bf16.main (cognitive 16) (inference/fp8_cast_bf16.py)
• generate.main (cognitive 16) (inference/generate.py)

Written by watchdog.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

deepseek-ai/DeepSeek-V3 was measured the same way every project in this corpus was: the same rubric, at a pinned commit, with the result published in full. Point a surveyor at a repository you know and see whether you agree with it.