Software & Apps

DeepSeek V4 reasoning models and agentic scaling trends

DeepSeek's V4-Pro-0813 model achieves a Terminal Bench 2.1 score of 87.9 using a DSpark speculative decoding module. The V4 series features massive parameter counts up to 1.6 trillion and utilizes the Fire-Flyer 2 cluster for efficient large-scale training and inference.

DeepSeek V4 reasoning models and agentic scaling trends

Model iterations and reasoning capability

DeepSeek’s V4 series models clearly outperform earlier iterations in reasoning and scale. The company released the 284-billion parameter DeepSeek-V4-Flash and the 1.6-trillion parameter DeepSeek-V4-Pro in April 2026. Full versions of these models arrived on July 31 and August 13. The DeepSeek-V4-Pro-0813 model version provides enhanced agentic capabilities through a DSpark speculative decoding module, which allows the model to achieve a Terminal Bench 2.1 score of 87.9 compared to the 72.1 score of the preview version. This release also introduces a reasoning_effort parameter that permits users to select between low, high, and max levels of deliberation. On the NL2Repo benchmark, the V4-Pro-0813 scored 61.5, while the V4-Flash scored 54.2. The V4.1-Flash model, released on September 10, uses a causal encoder-decoder architecture and beats Kimi K3 on cyber and coding benchmarks. The updated V4.1-Flash-1017, released on October 17, includes increased reinforcement learning training on coding and better compatibility with the DeepSeek Harness. The V4.1-Flash model is a natively multimodal system with lower memory usage than its predecessors. Previous models like V3.1, released in August 2025, utilized a hybrid architecture with thinking and non-thinking modes. Earlier, the V3.2-Exp model released in September 2025 used DeepSeek Sparse Attention to improve efficiency. However, users on the Google Play Store report that the model can get repetitive and frequently triggers "server busy" limits that prevent interaction for hours.

Model Version Parameter Count Context Window
DeepSeek-V4-Flash 284 Billion 1 Million Tokens
DeepSeek-V4-Pro 1.6 Trillion 1 Million Tokens

Infrastructure and compute efficiency

DeepSeek prioritizes training and inferencing efficiency to manage available compute and energy consumption. In September 2026, the company planned a data center in Inner Mongolia to host 160,000 Huawei Ascend 950DT chips. This massive deployment supports the high token demand seen in late September, such as the 1.3 trillion tokens processed on September 28. The company uses a modular hardware and software co-design for its Fire-Flyer 2 cluster. This cluster includes Nvidia GPUs using 200 Gbps interconnects and supports cross-zone tasks. The software relies on the 3FS distributed parallel file system, which uses Direct I/O and RDMA Read to handle random reads. The 3FS system uses Direct I/O, which means it does not cache data. The hfreduce library also runs asynchronously on the CPU to avoid blocking kernels on the GPU. DeepSeek also utilizes HaiScale Distributed Data Parallel to implement data, pipeline, tensor, and expert parallelism. The HAI Platform manages tasks such as scheduling, fault handling, and disaster recovery. The Fire-Flyer 2 cluster, which remained in operation as of 2025, contains 625 nodes, and each node contains 8 GPUs.

Agentic frameworks and usability

The DeepSeek Harness provides an environment to build and run AI agents through an "everything-is-a-plugin" design. As of August 2026, the company provides this harness as a developer preview. The Cordis framework powers this system. This allows developers to add or remove components while the software runs. This modular design enables developers to switch between third-party providers or alternative model configurations without changing the rest of the system. You should observe that the model struggles with context retention during long conversations. Users on the App Store report that the model often loses the thread of a conversation unless they restate the topic. The model also frequently reverts to Chinese even when users start chats in English. A user reported that the model incorrectly identified its creator as Anthropic or OpenAI. Furthermore, the model triggers a hard stop when requests brush against censorship guidelines, as it claims the request is "beyond its scope." The model also lacks a back key, which makes it difficult for creative writers to edit prompts. Does the company plan to address these language and identity inconsistencies in future V4 updates?