DeepSeek V4 Preview Takes Aim at AI Leaders With Million-Token Context and Lower Prices

Two Models Bring Scale and Lower Costs to the AI Race

DeepSeek released preview versions of DeepSeek-V4-Pro and DeepSeek-V4-Flash on April 24, 2026, pairing a context window of 1 million tokens with prices that put fresh pressure on OpenAI, Anthropic and Google. The flagship Pro model has 1.6 trillion parameters, while Flash has 284 billion, making the release the Chinese artificial intelligence laboratory's most ambitious attempt yet to compete with leading American systems.

The Hangzhou startup is pursuing two goals at once: stronger performance and cheaper access. V4-Pro reportedly comes close to Claude Opus 4.6 on a benchmark involving software repairs, while V4-Flash costs $0.28 per million output tokens. Both models are being released with accessible weights, the numerical settings that determine how a trained AI system responds, under the permissive MIT license.

The launch comes exactly one year after DeepSeek's R1 reasoning model triggered a market rout worth a trillion dollars by showing that advanced AI capabilities could be developed at a fraction of the costs associated with leading systems. V4 extends that challenge beyond price, with DeepSeek seeking to compete more directly on coding, mathematics and other demanding tasks.

DeepSeek still acknowledges limits. Its assessment places V4-Pro behind the most advanced competitors in general knowledge and reasoning, and a shortage of powerful computing equipment is restricting Pro service capacity. The release is a preview, arriving while the company is also reportedly seeking its first outside funding.

The central promise is straightforward: users could give the models much more material to work with, pay less for their responses and, with suitable hardware, run the released models themselves. The practical choices depend on which variant they use and how much computing equipment they can provide.

How a Trillion Parameters Becomes More Manageable

V4-Pro contains 1.6 trillion total parameters but activates about 49 billion during a given inference operation. Inference is the process of using an already trained model to produce an answer. V4-Flash has 284 billion total parameters, with about 13 billion active, and is aimed at applications where response speed matters and at deployment on consumer hardware.

Both use a Mixture-of-Experts architecture. Instead of involving every part of the model in every calculation, the system routes each input token to selected specialist networks. Tokens are the units of text a model processes, such as a word or part of a word. The approach gives the model access to a large pool of learned knowledge while limiting the amount of computation used for each step.

A dense model with 1.6 trillion parameters would involve the entire network and be prohibitively expensive to run. DeepSeek instead keeps V4-Pro's active parameter count roughly at the level of its previous V3.2 generation while expanding the pool of experts. The intended benefit is deeper specialization in coding, mathematics and multilingual tasks without a proportional increase in computing requirements.

V4-Pro is among the largest models with publicly accessible weights yet released. Flash takes the same broad approach at a smaller scale, offering developers a choice between the flagship's greater capacity and a model designed to place fewer demands on computing resources.

What the Million-Token Window Changes

Both V4 variants accept a context window of 1 million tokens, roughly enough text for 15 to 20 full novels in a single prompt. A context window is the amount of information a model can consider while preparing its response, including instructions and documents supplied by the user.

DeepSeek's Hybrid Attention Architecture combines Compressed Sparse Attention and Heavily Compressed Attention. Attention is the mechanism that helps a model connect relevant pieces of information. These methods aim to reduce the work and memory required when the material becomes very long.

DeepSeek claims 97% accuracy on retrieval tests involving a million tokens, compared with 84.2% for standard attention mechanisms. Such tests examine whether a model can locate a specific detail buried in a large volume of text, often called the needle-in-a-haystack problem. Conventional transformer attention can struggle to retrieve those details as the context grows.

The company's approach is intended to address that loss of retrieval quality. Its proposed uses include processing an entire software codebase, a large collection of legal documents or years of conversation history while retaining access to individual details.

Developers building retrieval-augmented generation systems could also benefit. Those systems usually find relevant information in an external collection before passing it to an AI model. A larger, more reliable context window could reduce or eliminate the need to divide documents into many pieces and maintain external vector databases, which organize information for searches based on meaning.

Where Pro Approaches Leading Models

V4-Pro's reported benchmark results, drawn from independent evaluations and DeepSeek's own reporting, place it close to several leading systems. On SWE-bench Verified, which tests whether a model can resolve real GitHub issues in projects with publicly available code, it scores approximately 80.6%. Claude Opus 4.6 scores 80.8%, putting the two close together on that measure and placing Pro above many previous records for models with accessible weights.

Mathematics is another reported strength. V4-Pro scores 95.2% on HMMT 2026 and 89.8% on IMOAnswerBench. On the latter, it is ahead of Claude Opus 4.6 but slightly behind GPT-5.4.

Competitive programming results also favor DeepSeek's flagship. Its reported Codeforces rating is 3206, compared with 3168 for GPT-5.4 and 3052 for Gemini 3.1 Pro. These evaluations focus on solving programming problems rather than the full range of work people might ask an assistant to perform.

DeepSeek's own assessment is more cautious about general knowledge and reasoning. The company places V4-Pro roughly three to six months behind Google's Gemini 3.1 Pro and other leading systems in development. In its assessment of V4-Pro-Max, which uses more reasoning tokens to work through problems, DeepSeek states:

Through the expansion of reasoning tokens, DeepSeek-V4-Pro-Max demonstrates superior performance relative to GPT-5.2 and Gemini-3.0-Pro on standard reasoning benchmarks. Nevertheless, its performance falls marginally short of GPT-5.4 and Gemini-3.1-Pro, suggesting a developmental trajectory that trails state-of-the-art frontier models by approximately 3 to 6 months.

Flash remains competitive despite its smaller scale. It scores 86.2% on MMLU-Pro, compared with Pro's 87.5%, and 91.6% on LiveCodeBench, compared with Pro's 93.5%. For developers whose work fits those capabilities, the smaller model offers a way to reduce operating costs while retaining much of the reported performance.

Prices Challenge Both Premium and Budget Rivals

V4-Flash is priced at $0.14 per million input tokens and $0.28 per million output tokens. Input tokens cover the material sent to the model; output tokens cover the text it produces. That separates the cost of asking a question from the cost of receiving a long answer.

Flash undercuts the listed prices of OpenAI's GPT-5.4 Nano, at $0.20 for input and $1.25 for output per million tokens, and Gemini 3.1 Flash-Lite, at $0.25 and $1.50 respectively. Among these budget options, DeepSeek's output price is particularly low.

V4-Pro costs $1.74 per million input tokens and $3.48 per million output tokens. Claude Opus 4.6 costs $5 for input and $25 for output, while GPT-5.4 costs $2.50 and $15. DeepSeek's flagship is positioned as providing comparable results on some benchmarks at costs roughly 50% to 80% lower than competing premium services.

The savings depend on the balance between incoming material and generated answers. For example, Pro's $3.48 output price is about 77% below GPT-5.4's $15, while its $1.74 input price is about 30% below GPT-5.4's $2.50. A task that generates lengthy answers therefore has a different cost comparison from one that mainly reads large documents.

DeepSeek previously launched R1 with API costs roughly 90% below comparable OpenAI services. Competitors subsequently adjusted prices and opened advanced models to users on free service plans. Industry analysts suggest V4 could create similar pressure across premium and budget AI services as developers compare the cost of accessing models through an application programming interface, or API.

Less Computation and Memory for Long Documents

DeepSeek attributes its pricing strategy to architectural efficiency. For contexts containing a million tokens, the company says V4-Pro uses only 27% of the floating-point operations required by V3.2 and 10% of its key-value cache size. V4-Flash uses 10% of the operations and 7% of the cache size compared with V3.2.

Floating-point operations, often shortened to FLOPs, are the numerical calculations performed by the hardware. The key-value cache is working memory that stores information from text already processed, helping the model avoid repeating work while generating an answer.

The reported figures translate into 73% less computation for Pro and 90% less for Flash in those long-context scenarios. The corresponding cache reductions are 90% and 93%. Reducing both calculation and memory demands helps explain how DeepSeek aims to support much longer inputs while keeping access prices low.

For a developer processing large collections of documents, these two costs matter together. A model must have enough memory to keep track of the material and enough computing power to produce an answer. DeepSeek's design targets both demands rather than relying solely on a larger context allowance.

Chinese Chips Add a Strategic Dimension

DeepSeek reportedly trained V4 using a combination of Huawei Ascend 910B AI accelerators and Cambricon MLU chips, with verification completed on Huawei's Ascend NPU platform. The reported infrastructure differs from the Nvidia equipment that underpins much development of leading AI models.

U.S. export restrictions limit China's access to Nvidia's most powerful chips. V4 therefore has significance beyond its benchmark scores: its development points to the possibility of building AI at this scale using Chinese silicon. If its performance claims hold up under independent scrutiny, Huawei and Cambricon would gain support as viable alternatives for large AI training projects.

Wei Sun, principal AI analyst at Counterpoint Research, said V4's ability to run natively on Chinese chips could have major consequences for Beijing's efforts to control its own AI infrastructure. Sun also described a wider effect:

This will ultimately speed up the global AI developments as well.

DeepSeek's reported use of domestic hardware connects its commercial ambitions with China's push to reduce dependence on American technology. It also points toward a more varied global supply chain for the equipment used to train and operate large AI systems.

What Running the Models Yourself Would Involve

DeepSeek says quantized versions of V4 can operate on consumer hardware. Quantization stores the model's numerical values with less precision, reducing the memory needed to hold them. The company suggests that V4-Flash can run on a single Nvidia RTX 5090 with 32GB of video memory using INT4 quantization, or on two RTX 4090 cards with 48GB in total using INT8 precision.

Those suggested configurations make Flash central to DeepSeek's accessibility pitch. Developers interested in keeping data under their own control, reducing response delays or making operating costs more predictable could consider running a model themselves rather than sending every request to a remote service.

Pro is aimed at a different scale of deployment. Its full weights occupy 865GB, and inference at full precision targets clusters of GPUs spread across multiple computing nodes. DeepSeek also indicates that Pro may support streaming deployment strategies, which would move model data through the system as needed.

The contrast with closed models is access and control. Their users depend on API access and the provider's data center infrastructure, while released weights permit developers to arrange their own deployment. DeepSeek's consumer hardware claims and its much larger Pro configuration describe different routes into that model of ownership.

The MIT license allows commercial use and modification without licensing fees. DeepSeek's broader release strategy also includes Apache 2.0 licensing, continuing its commitment to making model weights available as many Western laboratories keep their leading systems closed.

Competition, Funding and a Capacity Constraint

DeepSeek's announcement adds pressure within China as well as abroad. Since the R1 breakthrough, Alibaba's Qwen, ByteDance and Moonshot AI have accelerated model releases. MiniMax and Zhipu each fell approximately 8% in Hong Kong trading following the V4 announcement.

DeepSeek is also reportedly seeking external funding for the first time, targeting at least $300 million at a valuation exceeding $10 billion. Discussions involving Tencent and Alibaba could push the valuation toward $20 billion. Such an investment would mark a change from its earlier reliance solely on its parent, High-Flyer Capital Management, a hedge fund business.

DeepSeek has warned that Pro service capacity is currently limited by shortages of powerful computing equipment. It expects Pro prices to fall substantially after Huawei Ascend 950 supernodes are deployed at scale in the second half of 2026. No exact deployment date is given.

The launch also comes amid political conflict over AI development. The White House has recently accused China of intellectual property theft in AI on an industrial scale. Beijing's Foreign Ministry has rejected those allegations as groundless attacks on China's technological progress.

The immediate sequence and the next stated milestone are:

• One year before V4's launch: DeepSeek's R1 reasoning model triggered a market rout worth a trillion dollars and brought global attention to its lower costs.
• April 24, 2026: DeepSeek released preview versions of V4-Pro and V4-Flash.
• Following the announcement: MiniMax and Zhipu each fell approximately 8% in Hong Kong trading.
• Second half of 2026: DeepSeek expects deployment of Huawei Ascend 950 supernodes at scale, with substantially lower Pro pricing to follow.

V4 brings together several competitive pressures: prices below major rivals, reported coding and mathematics results close to leading systems, and deployment options spanning consumer hardware and large clusters. DeepSeek's own assessment of its reasoning gap and its warning about limited capacity remain central to understanding the preview as it stands on April 24.

The Bottom Line

• DeepSeek launched V4-Pro and V4-Flash previews on April 24, 2026, both with context windows of 1 million tokens.
• Pro has 1.6 trillion total parameters and 49 billion active; Flash has 284 billion total and 13 billion active.
• Both models are released under the MIT license, allowing commercial use and modification without licensing fees.
• Pro reportedly scores approximately 80.6% on SWE-bench Verified, close to Claude Opus 4.6's 80.8%.
• Flash costs $0.14 per million input tokens and $0.28 per million output tokens; Pro costs $1.74 and $3.48 respectively.
• DeepSeek reports large reductions in computation and cache memory compared with V3.2 for contexts containing a million tokens.
• V4 reportedly used Huawei Ascend 910B and Cambricon MLU chips for training, with verification on Huawei's Ascend NPU platform.
• DeepSeek proposes consumer GPU configurations for quantized Flash, while Pro's full weights occupy 865GB.
• The company places its strongest reasoning performance roughly three to six months behind the leading systems.
• Pro capacity is limited, with lower prices expected after Huawei Ascend 950 supernodes are deployed at scale in the second half of 2026.