DeepSeek V4 Preview: Architectural Innovations and Cost Restructuring Reshape LLM Applications

Introduction: The Era of Efficiency in LLM Competition

The evolution of large language models (LLMs) is undergoing a critical transition. Over the past two years, the primary focus of the industry was scaling parameter sizes. However, as models approach physical limits, a more fundamental question arises: how can LLMs do more with fewer resources under the practical constraints of high compute costs and latency sensitivity?

The release of the DeepSeek V4 preview lands precisely at this turning point. Instead of following the traditional path of stacking parameters, it targets the deepest pain points of the Transformer architecture: the computational bottleneck of attention mechanisms and the scaling困境 of context length. Simultaneously, its pricing strategy signals a historic drop in inference costs.

This article analyzes the architectural innovations, context processing mechanisms, and cost models introduced in this release.

1. Architectural Shift: “Lean and Precise” MoE Design

The total parameter count has doubled, while the activated parameters have only seen a minor increase. This contrast reveals a fundamental shift in V4’s design philosophy.

Traditional Mixture of Experts (MoE) implementations typically divide the model into specialized sub-networks, using a gating network to activate a subset during inference. While this expands model capacity, it also increases the selection pressure on the gating network, leading to imbalanced loads and knowledge fragmentation.

V4 introduces surgical improvements to the gating mechanism. The total parameter count has more than doubled from the previous generation’s 1.6 trillion, but the activated parameters have only increased from approximately 3.7 billion to 4.9 billion. This allows the gating network to select a highly precise subset of experts for each inference while significantly expanding the knowledge capacity of each expert.

This design directly benefits inference throughput. A larger knowledge pool combined with precise activation strategies reduces the computational overhead per request, allowing a single inference server to handle more concurrent requests.

2. Surgical Reconstruction of the Attention Mechanism

While MoE optimization focuses on resource allocation, V4’s changes to the attention mechanism are fundamentally structural.

Traditional Transformers face a structural bottleneck in long-context scenarios: as context length grows, the computational complexity of self-attention increases quadratically. V4 addresses this by decomposing attention calculations into two alternating modes, paired with a sliding window to retain local information:

Compressed Sparse Attention (CSA): Compresses the Key-Value (KV) cache of every 4 tokens into a single summary, creating a “directory-level” representation. Each Query only selects the top-K most relevant summaries for calculation, eliminating redundant processing of irrelevant historical context.

Heavily Compressed Attention (HCA): A more aggressive strategy that merges every 128 tokens into a summary but performs full computation on all summaries. This ensures the model captures global patterns over long distances.

Additionally, V4 retains a sliding window that continuously tracks recently generated raw tokens. Tokens within this window bypass compression and participate directly in attention calculations, ensuring high fidelity for local context details.

The results are significant. In a 1-million-token context scenario, the compute required per token drops to about 27% of the previous generation, and KV cache usage is reduced to 10%. The Flash version compresses these metrics further to 10% and 7%, respectively. Tasks that previously required multiple high-end GPUs with minute-level latency can now run on standard clusters in near real-time.

3. Million-Token Context: Balancing Capability and Engineering

Context window length dictates a model’s “working memory.” While the industry consensus previously capped practical usage at 128K to 256K tokens, V4 makes million-token context the default across its services, offering both Pro and Flash versions.

This engineering milestone mitigates the long-standing “context truncation” issue. In complex Agent tasks, models must retain user preferences, tool call intermediaries, and retrieved documents, easily breaching 256K limits. The million-token window provides ample buffer space.

However, longer context does not automatically guarantee flawless comprehension. The alternating use of CSA and HCA allows the model to scan historical information at varying granularities—using coarse summaries to locate relevant segments and finer compressed representations for details, while the sliding window ensures output coherence.

4. Restructured Pricing: Tailored for Agent Applications

Architectural innovations ultimately reflect in business costs. V4’s pricing strategy is particularly noteworthy.

For the Flash version, input costs 1 RMB per million tokens, dropping to 0.2 RMB upon a cache hit; output is 2 RMB per million tokens. For the Pro version, input is 12 RMB per million tokens, dropping to 1 RMB on a cache hit, with output at 24 RMB.

Compared to industry standards where output pricing for frontier models often exceeds tens of USD (over 200 RMB) per million tokens, V4-Flash’s pricing represents a cost reduction of over a hundredfold.

The tiered discount based on cache hits is critical. This design encourages placing static or semi-static content—such as system prompts, tool definitions, and conversation summaries—at the beginning of requests. For enterprise applications with high-frequency, multi-turn calls (like automated customer service or coding assistants), increasing the cache hit rate from 50% to 90% can reduce inference costs by nearly an order of magnitude.

5. Practical Recommendations for Developers

Based on V4’s features, developers should consider the following:

• Prioritize cache hit optimization: Place highly stable context (role settings, system instructions) at the start of the request and maintain a consistent request structure to maximize cache utilization.
• Select versions based on task profiles: Choose the Flash version for high-concurrency, cost-sensitive, and latency-critical tasks. Opt for the Pro version for complex reasoning, long-chain code generation, and multi-step tool calls.
• Re-evaluate long document processing: Million-token support allows many document analysis tasks previously dependent on Retrieval-Augmented Generation (RAG) to process entire documents directly, bypassing complex chunking and retrieval pipelines.

Conclusion: The Efficiency Revolution

The DeepSeek V4 preview marks a new phase in LLM competition: a race for architectural efficiency and cost engineering. The combination of compressed attention, precise parameter activation, and cache-hit pricing transforms million-token context from a technical showcase into an accessible engineering capability.

As inference costs drop dramatically, application paradigms will inevitably shift. Previously rejected ideas due to budget constraints are worth re-evaluating under this new architecture.