DeepSeek V4 Explained

DeepSeek V4 is DeepSeek's open-weights artificial intelligence (AI) model family, released under the Massachusetts Institute of Technology (MIT) license with a one-million-token context window. At Layer3Labs, we help businesses evaluate and integrate artificial intelligence systems into production workflows, where infrastructure control and inference unit economics dictate which models belong in enterprise architecture. The current production catalog centers on DeepSeek-V4.1-Flash under the Application Programming Interface (API) identifier deepseek-flash, alongside the larger DeepSeek-V4-Pro model.

DeepSeek released the original V4 preview in April 2026 with two sizes: a 1.6-trillion-parameter Pro model and a 284-billion-parameter Flash variant. By September 2026, DeepSeek retired the original Flash model and introduced V4.1-Flash, an updated 552-billion-parameter architecture that added native multimodal visual understanding and improved hardware efficiency. Both models provide production access through the OpenAI ChatCompletions schema and the Anthropic API format.

Evaluating DeepSeek V4 requires separating its technical design from its hosting options. Engineering teams can consume the models through a pay-as-you-go cloud API, access consumer chat interfaces, or download the model weights directly from Hugging Face to deploy on private infrastructure.

DeepSeek V4 Model Architecture and Lineup

DeepSeek V4 operates on a Mixture-of-Experts architecture designed to balance parameter scale against execution latency. Rather than activating every parameter for every incoming token, the models route requests through specialized subnetworks. The original April 2026 release introduced DeepSeek Sparse Attention (DSA) alongside token-wise compression, enabling the models to process 1,000,000 tokens of context without the memory scaling bottlenecks common to dense transformer architectures.

The flagship variant, DeepSeek-V4-Pro, contains 1.6 trillion total parameters while activating 49 billion parameters per token. In August 2026, DeepSeek updated the model to general availability with version DeepSeek-V4-Pro-0813, implementing Floating Point 4-bit (FP4) precision for MoE expert parameters and Floating Point 8-bit (FP8) precision across other model weights. The update also added native Responses API support for Codex developer integration.

The smaller branch underwent an architectural overhaul with the release of DeepSeek-V4.1-Flash in September 2026. This model implements a Causal Encoder-Decoder structure that activates 8 billion parameters on input and 16 billion parameters on output across a 552-billion-parameter total pool. DeepSeek claims that tests by multiple parties put V4.1-Flash ahead of V4-Pro on performance, cost, speed and total runtime.

• DeepSeek-V4-Pro: 1.6 trillion total parameters, 49 billion active parameters, 1M token context window, text and code input.
• DeepSeek-V4.1-Flash: 552 billion total parameters, 8 billion active input parameters, 16 billion active output parameters, native visual understanding.
• Engram conditional memory: 196 billion parameters sparsely accessed within V4.1-Flash to retain factual associations across long documents.
• Controllable reasoning: an integer setting from 1 to 100 within V4.1-Flash allowing developers to scale reasoning depth dynamically.
• Standardized interfaces: native protocol compatibility with OpenAI and Anthropic API request schemas.

Release Timeline and Version History

The DeepSeek V4 product line evolved through four dated releases between April and September 2026. The initial preview on April 24, 2026, established the one-million-token context baseline across the Pro and Flash sizes, publishing weights under the MIT license on Hugging Face. The release established competitive open-weights positioning against proprietary commercial models.

Rapid updates followed in August 2026. DeepSeek introduced the general availability build of DeepSeek-V4-Pro on August 13, optimizing agent execution paths and introducing FP4 quantization for expert parameters. Eight days later, on August 21, DeepSeek released DeepSeek-V4-Flash-Vision-Exp as an experimental multimodal testbed for image interpretation workloads.

The catalog shifted substantially on September 10, 2026, when DeepSeek launched DeepSeek-V4.1-Flash and announced the retirement of earlier Flash variants. The release post initially scheduled DeepSeek-V4-Pro for retirement on September 14, but DeepSeek reversed this decision in a change log update, citing user demand. DeepSeek committed to maintaining API access for V4-Pro without altering billing rates.

• April 24, 2026: DeepSeek-V4 preview debuts with V4-Pro (1.6T) and V4-Flash (284B), introducing 1M context windows and DSA.
• August 13, 2026: DeepSeek-V4-Pro-0813 reaches general availability with Codex integration and FP4 expert precision.
• August 21, 2026: DeepSeek-V4-Flash-Vision-Exp launches as an experimental visual reasoning release.
• September 10, 2026: DeepSeek-V4.1-Flash launches under model ID deepseek-flash, and V4-Flash and vision experimental models are retired.
• Before September 14, 2026: DeepSeek reverses the planned retirement of DeepSeek-V4-Pro, preserving ongoing API access for existing deployments.

Technical Specifications Across the Family

Comparing the technical specifications of the DeepSeek V4 family illustrates how DeepSeek compressed hardware requirements over time. While DeepSeek-V4-Pro relies on sheer scale with 1.6 trillion parameters, DeepSeek-V4.1-Flash achieves comparable benchmark outputs by routing tokens across 384 routed experts per MoE layer, activating six experts plus one shared expert per token. You can review detailed side-by-side trade-offs in our DeepSeek V4 Flash vs Pro comparison.

Hardware footprints differ sharply between the versions. DeepSeek reports that V4.1-Flash requires one-fourth the High Bandwidth Memory (HBM) and one-eighth the Solid-State Drive (SSD) storage capacity of the previous generation. The Key-Value (KV) cache for V4.1-Flash measures 890 bytes per token, permitting extended multi-turn conversations and document analysis inside constrained server memory pools.

Benchmark figures published on DeepSeek model cards reflect this architectural progression. On the MMLU-Pro benchmark, V4.1-Flash scores 74.1, exceeding V4-Pro at 73.5 and the retired V4-Flash at 68.3. On the HumanEval programming evaluation, V4.1-Flash records 79.4, compared to 76.8 for V4-Pro and 69.5 for V4-Flash. On the DeepSWE v1.1 software engineering test, V4.1-Flash reaches 74.2, substantially outperforming V4-Pro at 62.7.

• DeepSeek-V4-Pro: 1.6T total parameters, 49B active parameters, 1M context, 384K max output, text and code input only.
• DeepSeek-V4-Flash (Retired): 284B total parameters, 13B active parameters, 1M context, text and code input.
• DeepSeek-V4.1-Flash: 552B total parameters, 8B active input parameters, 16B active output parameters, 1M context, 384K max output, native vision.
• GSM8K math reasoning: V4.1-Flash achieves 93.0, V4-Pro achieves 92.6, and V4-Flash scores 90.8.
• Terminal-Bench 2.1 command-line evaluation: V4.1-Flash achieves 90.6, V4-Pro reaches 87.9, and V4-Flash scores 82.7.

Architectural Changes from DeepSeek V3

DeepSeek V4 introduces structural differences when contrasted with the preceding DeepSeek-V3 generation released on December 26, 2024. DeepSeek-V3 used a 671-billion-parameter MoE configuration that activated 37 billion parameters per token. The architecture was constrained to a 128,000-token context window, requiring external vector databases or retrieval systems for long-context ingestion. For historical context, consult our guide explaining DeepSeek V3 architecture.

The primary innovation in V4 is the expansion to a 1,000,000-token context window enabled by DeepSeek Sparse Attention. DSA compresses attention keys and values, allowing the network to retain context across hundreds of pages of technical documentation without the quadratic compute penalty of standard self-attention mechanisms. Our dedicated DeepSeek V4 vs DeepSeek V3 comparison examines these context mechanics in detail.

DeepSeek did not publish a direct, like-for-like benchmark comparison table matching V3 against V4 on the same model cards. The technical differentiation rests on parameter routing, native multimodal handling in V4.1-Flash, and the introduction of Engram conditional memory, rather than incremental score gains on older benchmark suites.

• Context window expansion: maximum token capacity increased from 128K tokens in V3 to 1M tokens in V4.
• Attention mechanisms: transition from standard dense attention matrices to DeepSeek Sparse Attention with token-wise compression.
• Parameter composition: V3 used 671B total and 37B active parameters, whereas V4.1-Flash uses 552B total with asymmetric 8B input and 16B output routing.
• Memory innovations: introduction of 196B-parameter Engram conditional memory in V4.1-Flash for factual document recall.

API Pricing and Commercial Deployment

DeepSeek operates its commercial API on a prepaid, pay-as-you-go balance model managed through platform.deepseek.com. There are no monthly platform subscription fees or recurring seat costs. Users fund their account balance and draw down tokens based on actual usage. Consumer interactions on chat.deepseek.com and the official mobile application remain free of charge.

Pricing is split between standard peak hours and discounted off-peak hours. DeepSeek defines the peak billing window as 01:00 to 04:00 UTC and 06:00 to 10:00 UTC, Monday through Friday. All token consumption occurring outside these specific windows receives an automatic 50 percent discount. DeepSeek adjusts these schedules periodically, so confirm the active windows on the official DeepSeek pricing documentation.

For teams running high-throughput production workloads, concurrency limits define system throughput. DeepSeek publishes a concurrency ceiling of 2,500 simultaneous requests for deepseek-flash and 500 simultaneous requests for deepseek-v4-pro. DeepSeek does not publish specific per-minute request (RPM) or token (TPM) caps. DeepSeek weights are also accessible through third-party platforms such as Amazon Web Services (AWS) Bedrock and OpenRouter at their own distinct rates, or teams can self-host the open weights via vLLM, SGLang, Docker Model Runner, Ollama, or llama.cpp.

• deepseek-flash (V4.1-Flash) input pricing: $0.006 per 1M tokens (cache hit, peak), $0.003 (cache hit, off-peak), $0.30 per 1M tokens (cache miss, peak), $0.15 (cache miss, off-peak).
• deepseek-flash output pricing: $1.20 per 1M tokens during peak hours, discounted to $0.60 per 1M tokens during off-peak windows.
• deepseek-v4-pro input pricing: $0.044 per 1M tokens (cache hit, peak), $0.022 (cache hit, off-peak), $1.32 per 1M tokens (cache miss, peak), $0.66 (cache miss, off-peak).
• deepseek-v4-pro output pricing: $3.96 per 1M tokens during peak hours, discounted to $1.98 per 1M tokens during off-peak hours.
• Concurrency allocations: 2,500 concurrent connections for deepseek-flash and 500 concurrent connections for deepseek-v4-pro.

Model ID Pinning and Catalog Volatility

Production systems integrating DeepSeek endpoints must manage catalog volatility through disciplined model identifier pinning. When DeepSeek introduced V4.1-Flash on September 10, 2026, DeepSeek automatically assigned the default deepseek-flash identifier to the new model. Legacy requests sent to deepseek-v4-flash or deepseek-v4-flash-vision-exp were temporarily redirected to V4.1-Flash without developer intervention.

The short-lived plan to deprecate DeepSeek-V4-Pro highlights why software teams should avoid dynamic aliases. DeepSeek announced on September 10 that deepseek-v4-pro requests would route to V4.1-Flash starting September 14. In response to user demand, DeepSeek reversed course within four days, keeping V4-Pro as an independent endpoint under version DeepSeek-V4-Pro-0813.

Software architectures requiring stable output distributions should pin explicit version identifiers in request payloads. Relying on rolling alias tags exposes systems to unannounced parameter updates, altered context handling, and changed reasoning behaviors. Tracking the DeepSeek change log prevents unexpected disruptions during infrastructure updates.

Cost Comparison Against Proprietary Frontier Models

The financial divergence between DeepSeek V4 and proprietary commercial alternatives is large for automated workflows. In September 2026, OpenAI released GPT-6 Astra, featuring a 1,050,000-token context window and a 128,000-token maximum output limit. OpenAI prices GPT-6 Astra at $10.00 per 1 million input tokens and $50.00 per 1 million output tokens on a closed-source infrastructure model. For the full rate card, see our GPT-6 Astra pricing overview.

Consider a production workload processing 2 million input tokens with cache misses and generating 500,000 output tokens monthly. On OpenAI GPT-6 Astra, that pipeline incurs $20.00 in input charges and $25.00 in output charges, totaling $45.00. Using deepseek-flash at peak rates, 2 million input tokens cost $0.60, and 500,000 output tokens cost $0.60, totaling $1.20. Running that same workload during off-peak hours drops the total to $0.60. DeepSeek reduces direct API inference spend by over 97 percent compared to GPT-6 Astra.

Open-weights licensing under MIT further differentiates the platforms. Closed models like GPT-6 Astra require sending data to OpenAI's infrastructure. ChatGPT access is sold by subscription: Plus at $20 monthly, Pro at $200 monthly, or Team at $25 per user monthly with a two-seat minimum. DeepSeek V4 weights can be downloaded, inspected, fine-tuned, and executed on private enterprise servers without per-seat software fees.

• Inference cost gap: deepseek-flash charges $0.30 input and $1.20 output (peak) versus GPT-6 Astra at $10.00 input and $50.00 output.
• Context dimensions: DeepSeek V4 provides a 1,000,000-token window with 384,000-token output, while GPT-6 Astra provides 1,050,000 tokens of input with 128,000 tokens of output.
• Licensing framework: DeepSeek provides open weights under the MIT license, while OpenAI maintains closed model weights accessible solely through proprietary APIs.
• Infrastructure choice: DeepSeek models support private hosting on company hardware, while proprietary frontier models require routing prompts to external multi-tenant cloud servers.

Appropriate Workloads and Data Restrictions

DeepSeek V4 suits high-volume data transformation, batch document extraction, code generation, and developer automation. Applications parsing extensive repositories or multi-file codebases benefit from the 1M context capacity at costs that do not penalize multi-pass prompt strategies. The ability to deploy open weights locally via vLLM or SGLang makes V4 a fit for private software tooling.

DeepSeek cloud API endpoints should not be used for regulated data pipelines. Organizations processing Protected Health Information (PHI) under the Health Insurance Portability and Accountability Act (HIPAA), sensitive defense intelligence, or strictly guarded customer financial data should avoid routing unencrypted traffic through overseas multi-tenant endpoints. Review our operational security assessment in the DeepSeek data privacy and security risks guide.

Regulated organizations requiring DeepSeek performance must deploy the MIT-licensed model weights internally within secure virtual private clouds (VPC) or on-premise hardware clusters. Running open weights on controlled infrastructure eliminates external telemetry risks while retaining the architectural advantages of the V4 family.

Next Steps for Production Adoption

Teams looking to implement DeepSeek V4 should construct an internal benchmark test using representative enterprise prompts across deepseek-flash. Evaluate token completion latency, context extraction accuracy across large prompts, and real-world cache hit ratios to project monthly operational budgets accurately. Compare outcomes using our DeepSeek V4 alternatives analysis.

Our recommendation would change if western cloud providers launch managed DeepSeek V4 endpoints with certified HIPAA and System and Organization Controls 2 (SOC 2) compliance guarantees. A substantial price reduction from closed-weights frontier providers closing the cost spread to commodity levels would also warrant reassessing deployment plans.

To begin deployment, register a testing account on the DeepSeek platform, fund a nominal prepaid balance, and pin deepseek-flash within a non-critical staging environment to assess DeepSeek V4 on your own prompts.

What you need to run DeepSeek V4 yourself

DeepSeek V4 is a frontier-scale Mixture-of-Experts model, so "running it yourself" is a real infrastructure decision — not something a single laptop or gaming GPU can do. Match the path below to how seriously you need to self-host. For most teams the API or rented GPUs are the right answer; buying hardware only pays off at steady, high volume or when your data can never leave your walls.

Once DeepSeek V4 is running, the fastest way to put it to work day to day is inside Cursor — point it at the model through OpenRouter as a custom model. And if you would rather run a model on one affordable box, see Best mini PCs for local AI and Local AI hardware calculator.

NVIDIA RTX 6000 Ada (48GB)

Power users and small clusters that want cards they own

View on Amazon →