DeepSeek introduced a new generation of artificial intelligence models, expanded its developer platform and published research aimed at improving computing efficiency during the first 11 months of 2026, while facing new restrictions on its use by U.S. federal civilian agencies.
The Chinese AI developer released its DeepSeek V4 family in April, followed by V4-Pro and V4.1-Flash updates designed for reasoning, software development and multimodal processing. It also expanded application programming interfaces for web search, document analysis, voice and automated tasks.
On October 20, the U.S. Department of Homeland Security published an order excluding DeepSeek products from federal civilian agency technology procurement and requiring their removal from covered government systems.
By November 21, DeepSeek’s developments reflected both technical advances in large AI models and growing questions about their use in sensitive government environments.
January: DeepSeek Publishes New AI Memory Research
DeepSeek began 2026 by researching more efficient ways to store and retrieve information during AI processing.
On January 12, the company published research titled Engram: Conditional Memory via Scalable Lookup.
The work examined an architecture that allows certain information to be retrieved through structured memory rather than requiring every operation to be handled by conventional neural network computation.
Such approaches could reduce the computing resources required for selected tasks and improve how models manage frequently encountered information.
According to DeepSeek’s official research archive, the study was one of several projects announced before the V4 model family arrived.
The findings were research results, not proof that the same improvements would appear in every production system.
January: DeepSeek-OCR 2 Investigates Visual Document Processing
On January 28, DeepSeek published DeepSeek-OCR 2: Visual Causal Flow, examining a new approach to interpreting visual information in documents.
Optical character recognition converts text in images or scanned pages into machine-readable information. More advanced document systems must also understand layout, reading order and relationships between different elements.
The research focused on improving how models process such visual structures rather than simply identifying individual characters.
These capabilities matter for AI applications that work with documents, forms, technical material, and scanned records.
February: DualPath Research Targets AI Inference Bottlenecks
DeepSeek published its DualPath research on February 25, addressing storage bandwidth limitations in agentic large language model inference.
AI agents often perform sequences of tasks involving reasoning, external tools, and repeated access to previous information.
These workloads can put significant pressure on memory systems and data storage because models must retrieve information repeatedly while generating new responses.
DualPath examined ways to reduce those bottlenecks and support more efficient processing.
The work reflected an important challenge for AI infrastructure: model performance depends not only on computing processors but also on how quickly information can move between storage, memory and processing systems.
April: DeepSeek V4-Pro and V4-Flash Launch in Preview
DeepSeek introduced its V4 model family on April 24, releasing previews of DeepSeek-V4-Pro and DeepSeek-V4-Flash.
According to the company’s official release announcement, V4-Pro contained 1.6 trillion total parameters, with 49 billion activated during relevant processing operations.
V4-Flash was a smaller model with 284 billion total parameters and 13 billion active parameters.
Both models used a mixture-of-experts architecture, which activates selected portions of a model instead of using all parameters for every token.
DeepSeek said the models supported a context window of up to one million tokens, allowing them to process substantially longer inputs than conventional shorter-context systems.
The company also introduced updated attention mechanisms intended to reduce memory and computing requirements.
DeepSeek described improvements in reasoning, software development and agentic coding, although the performance figures came from its published evaluations.
The company made the models available through its web interface and API, with separate modes for demanding reasoning tasks and faster interactions.
June: Research Explains the V4 Architecture
On June 24, DeepSeek published research describing V4 as an architecture designed for highly efficient processing of million-token contexts.
The work examined how attention mechanisms and information compression could reduce the resources required to process long sequences.
Large context windows are useful for applications involving extensive documents, software repositories and tasks that require information from earlier interactions.
However, a model’s ability to accept a million tokens does not automatically establish that it can retrieve or reason about every part of that input with equal accuracy.
DeepSeek’s research focused on improving the efficiency of long-context processing while maintaining performance across demanding tasks.
July and August: DeepSeek Expands V4 Availability
DeepSeek released an updated V4-Flash API in public beta on July 31.
The company said the update improved performance on agentic software development benchmarks and added support for additional API workflows.
DeepSeek also confirmed that the update applied specifically to V4-Flash. That release did not change its V4-Pro service and main application.
On August 13, DeepSeek announced the general availability of V4-Pro across its application, website and API.
In its official announcement, the company introduced three reasoning-effort settings: low, high and max.
These settings let developers adjust processing effort based on task complexity.
DeepSeek also introduced peak and off-peak API pricing, with off-peak rates set at half the corresponding peak rates from August 16.
The company positioned the update as part of its efforts to balance model capability, inference cost and demand for computing resources.
TechAmerica.ai previously examined the DeepSeek V4-Pro architecture and performance, including the distinction between model size and active parameters.
August: DeepSeek Introduces Experimental Vision Model
On August 21, DeepSeek released V4-Flash-Vision-Exp, an experimental model designed to process visual information alongside text.
The model expanded V4-Flash to handle tasks that require understanding images and other visual inputs.
According to DeepSeek’s API announcement, the experimental version also supported agent workflows involving visual understanding.
The company reported improvements on selected evaluations compared with the text-only V4-Flash model.
These results were based on specific benchmark conditions and did not guarantee the same improvements across every real-world application.
September: DeepSeek V4.1-Flash Introduces New Architecture
DeepSeek announced V4.1-Flash on September 10, introducing a model with native multimodal visual understanding and an updated computing architecture.
The company described a 552-billion-parameter mixture-of-experts model using an asymmetric encoder-decoder design.
According to DeepSeek’s official announcement, the architecture activated about 8 billion parameters for input processing and 16 billion for output generation.
DeepSeek also reported that the model required less memory for its key-value cache, which stores information used during token processing.
The company said the reduced cache requirements could help lower the operating costs of lengthy AI interactions.
DeepSeek made V4.1-Flash available through the DeepSeek API with the model identifier deepseek-flash.
Earlier V4-Flash API names were temporarily redirected to the updated model for compatibility.
DeepSeek initially announced plans to phase out V4-Pro API access, but its updated change log said the V4-Pro service would continue due to user demand.
October: DeepSeek Improves Coding Agents and Opens Search APIs
On October 17, DeepSeek released V4.1-Flash-1017, an updated model designed to improve software development and multistep agent tasks.
The company described changes intended to reduce task completion time, shorten responses and improve performance on selected coding benchmarks.
In its official announcement, DeepSeek also opened access to its Search and Fetch APIs.
The search services let developers retrieve information from the web and use it in AI-powered applications.
DeepSeek said its internal search system handled hundreds of millions of requests daily and indexed a large collection of web pages.
Those scale figures were company claims, not independently audited measurements.
The company also continued publishing open-source development tools, including updates to software for connecting model outputs with developer applications.
TechAmerica.ai previously covered DeepSeek Harness and its extensible AI agent framework, which allows supported tools and capabilities to be managed within automated workflows.
October: U.S. Government Orders DeepSeek Removal From Civilian Systems
DeepSeek faced a major U.S. regulatory development on October 20, when the Department of Homeland Security published an order concerning its products in federal civilian agencies.
The Federal Register notice excluded DeepSeek products and services from procurement by covered federal civilian agencies.
It also ordered the removal of DeepSeek products and certain third-party products developed using DeepSeek implementations from civilian agency information technology systems.
The action followed a review under the Federal Acquisition Supply Chain Security Act and the applicable federal procurement regulations.
The order applied to the specified federal civilian technology environment. It did not establish a general prohibition on every private American citizen or company using DeepSeek.
TechAmerica.ai previously reported on the U.S. order removing DeepSeek from federal civilian agencies and its implications for government technology.
The decision demonstrated the growing importance of supply-chain security and procurement controls in federal AI adoption.
November: DeepSeek Opens More AI Application Capabilities to Developers
DeepSeek announced a substantial expansion of its developer platform on November 13.
The company introduced API access to features used in its main application, including web search, document input, URL reading, voice processing and safety checks.
According to DeepSeek’s official API announcement, developers could enable several functions through existing Messages, Responses, and Chat Completions interfaces.
Search functionality supports web information retrieval, while URL reading lets a model retrieve content from links supplied during a conversation.
File input supports processing documents and images, adding relevant information to the model’s available context.
DeepSeek also introduced access to optical character recognition, automatic speech recognition and text-to-speech services through separate APIs.
The company emphasized that developers offering public-facing applications remained responsible for content safety, applicable regulatory registration and AI-generated content labeling.
On November 18, DeepSeek updated its Search and Fetch billing arrangements, removing separate charges for certain server-side tool calls while continuing to charge for the model tokens generated from retrieved information.
DeepSeek’s Financial Disclosures Remain Limited
DeepSeek also attracted attention in 2026 over reports about financing, business valuation, and ownership of its technology operations.
However, the available official research and API publications do not establish a verified completed funding round or a publicly confirmed valuation for the period covered by this report.
Figures circulating in unofficial reports should therefore not be treated as completed investments without supporting company or regulatory documentation.
Likewise, publicly available sources reviewed for this report did not establish a comprehensive accounting of DeepSeek’s revenue, profitability or infrastructure spending.
DeepSeek’s 2026 Developments Combine AI Efficiency and Regulatory Pressure
Between January and November 21, DeepSeek’s announcements showed continued investment in large language model architecture, efficient inference and developer-facing tools.
The V4 family brought a million-token context capability, while later updates expanded multimodal processing and software development performance.
The company’s research also addressed the storage and memory requirements that can make large AI systems expensive to operate.
At the same time, the October U.S. federal order demonstrated that model capability and low operating costs are not the only considerations for government adoption.
Security requirements, procurement restrictions and the handling of sensitive information can determine whether AI technology is permitted in public-sector systems.
By November 21, DeepSeek had expanded its models and application interfaces while confronting significant government restrictions in the United States. The year’s developments illustrated the connection between advances in AI engineering and the policy decisions governing where those systems can be deployed.








