The model is explicitly identified as a distilled version of DeepSeek-R1 using the Llama-3.1-8B-Instruct base architecture. The training methodology is documented in the DeepSeek-R1 technical report, detailing a multi-stage pipeline involving cold-start data, reinforcement learning (GRPO), and rejection sampling. While the distillation process (SFT on 800k samples) is described, the specific architectural modifications beyond the standard Llama-3.1 dense transformer structure are minimal, though the integration of the reasoning 'thinking' phase is well-documented.

DeepSeek discloses that the 8B variant was fine-tuned on 800,000 samples generated by the larger DeepSeek-R1 model. However, the composition of the original pre-training data for the underlying Llama-3.1 base is inherited from Meta's disclosures, and the specific breakdown of the 800k reasoning samples (e.g., math vs. code vs. logic proportions) is not provided in detail. The 'cold-start' data used for the teacher model is described as 'thousands' of samples but lacks a public source or comprehensive breakdown.

The tokenizer is publicly available via the Hugging Face repository and is based on the Llama-3.1 tokenizer with a vocabulary size of 128,256 tokens. Documentation confirms the use of Byte-Pair Encoding (BPE). While there were initial community reports of minor config.json mismatches regarding embedding sizes in some distilled variants, the functional tokenizer is verifiable and matches the claimed language support and reasoning token (<think>) integration.

The model is a dense 8B parameter architecture, unlike its MoE teacher. The parameter count is clearly stated and consistent across official documentation, Hugging Face, and third-party platforms like Ollama. There is no ambiguity regarding active vs. total parameters as it does not use a sparse Mixture-of-Experts design.

While DeepSeek provides high-level compute figures for the teacher model (2,048 NVIDIA H800 GPUs for two months), specific compute metrics for the distillation of the 8B variant itself are vague. There is no detailed disclosure of the GPU hours, carbon footprint, or specific hardware used for the 800k-sample SFT phase of this specific 8B model.

Evaluation results for AIME, MATH-500, and LiveCodeBench are provided in the technical report. However, the exact evaluation code and full prompt sets used for the distilled variants are not as thoroughly documented as the main R1 model. Third-party verification has shown significant variance in results, and the reliance on synthetic data for distillation introduces complexities in verifying the 'cleanliness' of the evaluation process.

The model consistently identifies itself as a DeepSeek-distilled version of Llama. It maintains a clear versioning identity (e.g., the 0528 update) and does not exhibit the identity confusion common in some fine-tuned models that claim to be GPT-4 or other competitors. It is transparent about its nature as a reasoning-focused model.

The model weights are released under the MIT License, which is highly permissive. However, because it is built on Llama-3.1-8B, users must also adhere to the Meta Llama 3.1 Community License Agreement. This dual-licensing creates a slight complexity for commercial users, although DeepSeek's own contributions are clearly MIT-licensed.

Hardware requirements are well-documented by both the provider and the community. VRAM requirements for FP16 (~16-20GB) and various quantization levels (4-bit, 8-bit) are widely available. Documentation from partners like NVIDIA and community tools like Ollama provide clear guidance on running the model on consumer hardware.

DeepSeek has released versioned updates (e.g., the 0528 update), but the changelogs are relatively high-level and lack granular detail on specific weight changes or performance drift across all benchmarks. While semantic versioning is partially applied, the history of changes between the initial January release and subsequent updates is not comprehensively archived in a public changelog.