Leiphone Report Details DeepSeek's DSec Sandbox and Liang Wenfeng's 11 Authored Papers

The 31-page paper says the system can automatically run 3 million sandbox environments per day, with a single production unit spanning 160 CPU nodes, according to Leiphone. Strict permission isolation is used to prevent AI cheating and to allow precise reproduction of every step of environment feedback, the report says. Leiphone frames this work as an engineering push in the current reinforcement-learning and agent era, when sandboxes serve both as training grounds and as safeguards for agents that can behave unpredictably.

The report says Liang has rarely appeared as first or corresponding author on flagship model reports such as V3.2, V4 or V4.1-Flash, which carry hundreds of authors, but has repeatedly signed papers on foundational technologies such as MLA attention, MoE routing, mHC topological manifolds, DSpark inference operators and the DSec agent sandbox. Over three years, Leiphone counts 11 papers signed by Liang: DeepSeek LLM: Scaling Open-Source Language Models with Longtermism; DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models; DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model; DeepSeek-Coder-V2: Breaking the Barrier of Closed-Source Models in Code Intelligence; DeepSeek-V3 Technical Report; DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning; Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention; Insights into DeepSeek-V3: Scaling Challenges and Reflections on Hardware for AI Architectures; mHC: Manifold-Constrained Hyper-Connections; DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation; and DSec: An Efficient Sandbox Infrastructure for Large-Scale Agent Training.

Leiphone links the first group of papers to early 2024, when the industry was locked in a dense-model compute race. It says GPT-4-class models required more than 10,000 high-end GPUs and a single training run cost more than $63 million in materials, while companies such as Stability AI and Inflection AI struggled with rising costs. The report says DeepSeek LLM, published on Jan. 5, 2024, trained 7B and 67B versions on 2 trillion high-quality Chinese-English tokens; the 67B version surpassed Llama-2 70B on code, math and reasoning benchmarks and beat GPT-3.5 in open dialogue tests, according to Leiphone.

The report then describes DeepSeekMoE and DeepSeek-V2 as a reconstruction of the traditional Transformer. It says DeepSeekMoE's fine-grained expert routing and shared-expert isolation cut training overhead by 42.5%, while the MLA mechanism compressed KV Cache volume by 93.3%. According to Leiphone, DeepSeek-V2 delivered GPT-4-level performance at a fraction of the cost and helped trigger an API price war in China.

In June 2024, the report says, DeepSeek-Coder-V2 injected 6 trillion high-quality code and math tokens, expanded supported programming languages from 86 to 338 and raised the context window to 128K. It says the model surpassed GPT-4 Turbo and Claude 3 Opus on benchmarks including HumanEval, Codeforces and LiveCodeBench, while pricing input at $0.14 per million tokens and output at $0.28 per million tokens, compared with $10 and $30 for GPT-4 Turbo.

At the end of 2024, Leiphone says, DeepSeek trained DeepSeek-V3 in less than two months using 2,048 H800 GPUs. The model had 671 billion parameters and a stated total cost of $5.6 million, a figure the report says drew attention in Silicon Valley even after analysts such as SemiAnalysis noted it did not include earlier research losses, equipment depreciation or a broader $1.3 billion hardware bill. The V3 paper attributed efficiency gains to an auxiliary-loss-free expert load-balancing strategy, FP8 training precision and overlapping computation with data transfer.

The report says DeepSeek-R1, released around the 2025 Spring Festival, used the GRPO algorithm and completed its reinforcement-learning stage for $294,000. It says R1-Zero started from zero without human-labeled data and reached a level close to top closed-source reasoning models through pure reinforcement learning. According to Leiphone, the work later appeared on the cover of Nature and was covered by MIT Technology Review, The New York Times, The Washington Post, CNBC and the Financial Times as evidence that U.S. restrictions could not stop Chinese AI progress.

Leiphone also says Liang's later work moved further down the stack, with Native Sparse Attention and Insights into DeepSeek-V3 addressing long-context training and inference costs at the level of chip memory access. The report presents the 11 papers as a technical campaign aimed at compute efficiency, foundational architecture and agent training rather than at flagship model publicity.