DeepSeek-V3.2

DeepSeek-V3.2 is extensively documented through a technical report and model cards. It explicitly builds on the DeepSeek-V3/V3.1-Terminus architecture, utilizing a Mixture-of-Experts (MoE) framework with 671B total and 37B active parameters. Key architectural innovations like DeepSeek Sparse Attention (DSA) and Multi-head Latent Attention (MLA) are detailed with mathematical formulations and diagrams in the official paper. The training methodology, including the use of Multi-Token Prediction (MTP) and an auxiliary-loss-free load balancing mechanism, is publicly described.

While the total token count (14.8 trillion) and general categories (web pages, e-books, code, math) are disclosed, the specific proportions of the dataset are not provided in detail. The documentation mentions 'enhancing the ratio' of math and code but lacks a granular percentage breakdown. Information on the '128K long context extension data' is described as 'aligned' with previous versions but lacks specific source disclosure. The model uses a 'Large-Scale Agentic Task Synthesis Pipeline' for post-training, which is documented as a methodology, but the underlying raw data sources remain largely opaque.

The model uses the same tokenizer as DeepSeek-V3, which is publicly accessible via Hugging Face (LlamaTokenizerFast). The vocabulary size and special tokens (e.g., for tool use and reasoning blocks) are clearly defined in the tokenizer_config.json. Documentation explicitly notes the stability of the tokenizer across versions V3 to V3.2, and the 128K context window alignment is verified through public configuration files.

DeepSeek provides exemplary transparency regarding its MoE architecture. It clearly distinguishes between total parameters (671B) and active parameters (37B). The breakdown of experts (256 routed experts + 1 shared expert) is explicitly stated. This level of detail prevents the common 'parameter inflation' seen in other MoE models and allows for accurate computational modeling by third parties.

The technical report provides specific compute metrics, stating the use of 2,048 NVIDIA H800 GPUs over approximately two months. Total training compute is disclosed as 2.788 million GPU hours, with a cost estimate of ~$5.576M based on rental rates. While it lacks a formal third-party carbon audit, the provided energy consumption data (180K GPU hours per trillion tokens) allows for independent environmental impact calculations.

DeepSeek discloses scores across a wide array of standard benchmarks (MMLU-Pro, AIME 2025, SWE-Bench) and provides some evaluation scripts and chat templates on GitHub. However, the full evaluation pipeline and exact few-shot prompts for all benchmarks are not as comprehensively documented as the architecture. There is a noted discrepancy between internal math scores (~89% AIME) and some independent estimates, though the model's performance is generally verifiable through public leaderboards like LMSYS.

The model consistently identifies itself as a DeepSeek-developed AI and maintains version awareness (V3.2). It is transparent about its 'thinking' vs 'non-thinking' modes and the specific limitations of variants like 'Speciale' (which lacks tool-calling). There are no documented instances of the model claiming to be a competitor's product or denying its nature as an AI.

The model is released under the MIT License, which is one of the most permissive and transparent open-source licenses available. The license terms are clearly stated on GitHub and Hugging Face, explicitly permitting commercial use, modification, and redistribution without the restrictive 'acceptable use' clauses often found in other 'open' weights models.

Hardware requirements are well-documented for various deployment scenarios. Official and community guides (vLLM, SGLang) provide specific VRAM requirements for FP16 (~1.5TB) and 4-bit quantization (~350-400GB). The impact of context length on memory scaling is addressed, and the documentation provides clear guidance on multi-GPU configurations (e.g., 8x or 16x 80GB GPUs) required for efficient inference.

DeepSeek maintains a clear versioning history (V3 -> V3.1 -> V3.2) with associated technical updates for each. While it lacks a formal, real-time 'drift' dashboard, the release of specific checkpoints (e.g., 0324, 0528) and the use of semantic-style versioning allow developers to track changes. The transition from experimental (Exp) to stable releases is documented, though detailed changelogs for minor weight updates could be more granular.