DeepSeek released a preview of its V4 model as the competition in artificial intelligence intensifies. This release follows the success of the R1 reasoning model, which achieved high performance with less computing power than its American counterparts. The company trained R1 with lower-capacity Nvidia chips using only two months of work and less than $6 million. This ability to match or exceed the benchmarks of leading large language models using cheaper hardware challenged the assumption that massive computing infrastructure is the only way to build frontier AI. US export controls on advanced chips like the Nvidia H100 and H200 intended to slow Chinese progress. These restrictions forced Chinese engineers to prioritize algorithmic efficiency. DeepSeek’s V4 development follows this pattern of using smart architecture to overcome hardware limits.
The current landscape of AI development shows that hardware availability is not the only factor in model performance. While US labs spend billions on massive computing clusters, DeepSeek focuses on making models that require less raw compute. This strategy lowers the barrier for deployment and changes the economic math for AI providers. The company’s ability to produce competitive models with limited access to top-tier chips suggests that innovation can thrive under constraint. As developers in China adopt domestic alternatives, the reliance on American silicon decreases.
| Model Version | Parameters (Total) | Active Parameters (Prefill) | Context Window |
|---|---|---|---|
| DeepSeek V4.1 Flash | 552B | 8B | 1,000,000 tokens |
| DeepSeek V4 Pro | 1.6 Trillion | Variable | 1,000,000 tokens |
Architecture of the V4.1 Flash Model
DeepSeek-V4.1-Flash is the smallest model in the new architecture family. It uses a Causal Encoder-Decoder (CED) design. This design consists of a 40-layer Transformer. The architecture splits the layers into a 20-layer causal encoder and a 20-layer decoder. In this setup, the decoder’s global KV cache is projected from the final hidden states of the encoder instead of being rebuilt layer by layer. This method helps reduce the computational load. The model uses a Mixture-of-Experts (MoE) architecture. It has 552B backbone parameters. During the prefill stage, the model only activates 8B parameters per token. During decoding, it activates 16B parameters per token.
The model also uses Engram conditional memory technology. This technology separates static pattern retrieval from dynamic reasoning. This separation helps the model manage long contexts more efficiently. A 196B-parameter conditional memory module is accessed through token-based lookup. This system helps maintain coherence when processing extremely long prompts. The V4.1 Flash model also includes native multimodal visual understanding. It can process images and text in a single pass. The model was trained on 45T tokens. The context was extended to 1 million tokens at the 34T mark.
Inference Economics and KV Cache Compression
Inference costs involve the financial and computational expenses of running a model to generate outputs. DeepSeek V4.1 Flash achieves lower costs through specialized memory management. The model uses Compressed Sparse Attention 2 (CSA2) and FP4 KV caching in E2M1 format. These technologies shrink the global KV cache to 890 bytes per token. This is one quarter of the footprint of the previous V4-Flash generation. The SWA Bounded Replay technique further reduces the persistent KV cache to one eighth of the previous Flash version.
Reducing the KV cache footprint has direct implications for agentic tasks. In agentic workflows, models frequently re-read the same context. High cache-hit charges often account for a large part of the cost for these tasks. Because V4.1 Flash compresses the cache so effectively, it can process large contexts for much less money. The model also supports three different thinking effort levels: low, high, and max. Users can select low for simple tasks and max for complex reasoning. The ability to scale reasoning effort allows for better control over the cost-to-accuracy ratio.
| Metric | V4.1 Flash (Peak) | V4.1 Flash (Off-Peak) | V4 Pro (Peak) |
|---|---|---|---|
| Input per 1M tokens | $0.30 | $0.15 | $1.32 |
| Output per 1M tokens | $1.20 | $0.60 | $3.96 |
Performance Against Frontier Models
DeepSeek V4.1 Flash shows competitive performance on several agentic benchmarks. On the DeepSWE v1.1 benchmark, the model scores 74.2. This is higher than the 73.0 score of OpenAI’s GPT-5.6 Sol. In the AutomationBench test, V4.1 Flash scores 54.8, while Sol scores 45.8. The model also scores 31.8 on the Agents’ Last Exam, which is higher than the 26.7 score from Sol. In CyberGym, V4.1 Flash achieves 88.1, beating the 84.5 score of Sol.
The model does not win in every category. Claude Opus-5.0 leads in long-horizon terminal work. Specifically, Opus-5.0 scores 43.3 on Terminal-Bench 3.0, while V4.1 Flash scores 30.0. In the SEC-Bench Pro test, Sol achieves 74.3, which is higher than the 62.8 score from V4.1 Flash. The model also trails in the GPQA Diamond benchmark, where Sol scores 94.1 compared to the 90.9 from V4.1 Flash. Most wins for V4.1 Flash are in short-horizon, high-volume agent tasks.
| Benchmark | DeepSeek V4.1 Flash | GPT-5.6 Sol | Claude Opus-5.0 |
|---|---|---|---|
| DeepSWE v1.1 | 74.2 | 73.0 | N/A |
| AutomationBench | 54.8 | 45.8 | N/A |
| Agents’ Last Exam | 31.8 | 26.7 | N/A |
| CyberGym | 88.1 | 84.5 | N/A |
| SEC-Bench Pro | 62.8 | 74.3 | N/A |
Domestic Hardware and the Huawei Surge
The development of V4 was influenced by the availability of domestic Chinese hardware. DeepSeek optimized V4 to run on Huawei’s Ascend 950 chips. This design choice changed the procurement behavior of major Chinese tech firms. ByteDance, Alibaba, and Tencent have all placed significant orders for Huawei’s Ascend 950 processors. ByteDance committed to multi-billion-dollar orders for the Ascend 950PR. Combined, the orders from ByteDance, Alibaba, and Tencent exceed 500,000 units. This amount is approximately 67 percent of Huawei’s planned production of 750,000 units for 2026.
Huawei began mass production of the 950PR in April 2026. Full-scale shipments are expected in the second half of the year. The Ascend 950PR is the only domestic chip that supports processing AI calculations in a highly compressed format. This technical compatibility makes it a logical choice for companies deploying V4 at scale. The scramble for these chips is intensified by US restrictions on advanced chipmaking equipment. Global memory chip supplies for AI hardware are also sold out through 2026. Will the demand for Huawei silicon eventually outpace their production capacity?
Licensing and Commercial Use Rules
DeepSeek provides its models under different license types for code and model weights. The source code uses the MIT license. The model weights use a custom license based on the OpenRAIL framework. This custom license focuses on adaptability and openness while requiring responsible use. It allows users to use, reproduce, and distribute the model for any lawful purpose. This includes direct deployment, fine-tuning, and creating proprietary products.
The license includes specific use-based restrictions. Users cannot use the model for military applications. The license also prohibits generating false or harmful content. Users must also avoid violating personal rights. Any derivative works must include these same use-based restrictions. DeepSeek does not charge fees for the use of these open-source models. Developers can build applications and generate revenue without sharing profits with DeepSeek. Because the license is not a copyleft license, developers can choose whether to open source their own derivative models. You should review the specific use restrictions before integrating these weights into a commercial application.
Pricing Structures and Model Routing
DeepSeek changed its API pricing and model availability in September 2026. On September 10, 2026, the company released V4.1 Flash. This release led to the retirement of V4 Flash and V4 Flash Vision Exp. Requests to those old names now route to V4.1 Flash. On September 14, 2026, DeepSeek will begin routing V4 Pro requests to V4.1 Flash. This change happens at 12:00 Beijing Time, which is 04:00 UTC. V4 Pro requests will then be billed at the lower V4.1 Flash rates.
The pricing for V4.1 Flash depends on the time of day. Peak pricing applies during certain hours, while off-peak pricing is 50 percent of the peak rate. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. For users in Australia, this means peak hours fall between 11:00 to 14:00 and 16:00 to 20:00 AEST.
| Model | Input Type | Input Price (per 1M) | Output Price (per 1M) |
|---|---|---|---|
| V4.1 Flash | Peak | $0.30 | $1.20 |
| V4.1 Flash | Off-Peak | $0.15 | $0.60 |
| V4 Pro | Peak | $1.32 | $3.96 |
| GPT-5.6 Sol | Peak | $2.00 | $10.00 |
Deployment Strategies for AI Builders
Building on DeepSeek models requires understanding the specific strengths of the V4 family. For high-volume tasks like document summarization or classification, V4.1 Flash is a cost-effective option. These tasks do not require frontier intelligence but do require reliable inference. For agentic workflows, the low output token cost of V4.1 Flash provides a massive economic advantage. Most agentic loops generate large amounts of text, so output pricing is a primary driver of total cost.
The open-weight nature of V4 allows for local deployment. Organizations with existing GPU infrastructure can download the weights and run the model on their own hardware. This provides a way to maintain data sovereignty. You can run the model behind a private firewall to keep data within your own network. For users with high-volume automation needs, scheduling batch jobs during off-peak hours will reduce costs by half. This works because the API supports flexible scheduling based on the user’s workload. Use the deepseek-flash model name to ensure you access the latest updates to the V4.1 Flash tier.
