DeepSeek V4 is a 2026 Mixture-of-Experts family with a hybrid long-context attention design. Pro publishes 1.6T total / 49B activated parameters; Flash publishes 284B / 13B. This page explains what those numbers mean, how CSA and HCA target million-token efficiency, what the paper actually reports, and where application claims must stop.
Architecture summary
The official model card summarizes architecture and optimization changes at the family level. These mechanisms are internal to the model; they are not switches exposed by this site.
Compressed Sparse Attention selects and compresses relevant long-range information, while Heavily Compressed Attention provides an aggressively compressed path. DeepSeek combines them to avoid paying uniform dense-attention cost across a million-token sequence.
mHC augments conventional residual connections. The publisher says it improves signal-propagation stability while preserving expressivity. This is an architecture claim from the report, not an independent measurement by this website.
The model card says the Muon optimizer is used to improve convergence and training stability. Optimizer choice describes how weights were trained; it does not determine a user-visible API parameter or guarantee a benchmark result.
Model specifications
Pro and Flash share the V4 family architecture and published context, but they are distinct weight sets and API model IDs. Benchmark or parameter facts for one must not be silently assigned to the other.
| Field | V4 Pro | V4 Flash |
|---|---|---|
| Model ID | deepseek-v4-pro | deepseek-v4-flash |
| Architecture family | Mixture-of-Experts (MoE) | Mixture-of-Experts (MoE) |
| Total parameters | 1.6T | 284B |
| Activated parameters | 49B per token | 13B per token |
| Published context | 1M tokens | 1M tokens |
| Published max output | 384K tokens | 384K tokens |
| Instruct precision | FP4 experts + FP8 mixed | FP4 experts + FP8 mixed |
| Modes | Non-think, Think High, Think Max | Non-think, Think High, Think Max |
MoE parameters
A Mixture-of-Experts model contains many expert parameter blocks. A router selects only part of that expert capacity for each token. The 1.6T Pro figure describes the complete weight set; 49B describes the published activated subset. Flash’s equivalent figures are 284B and 13B.
Activated count is still not a complete latency or hardware estimate. Serving also depends on weight precision, memory bandwidth, expert placement, communication, batching, attention/KV cache, context, kernels, and provider infrastructure. Local storage must account for the full selected weight artifact even when a token activates fewer parameters.
Plan local deployment from the exact weightsThe bar is intentionally illustrative, not a scale-accurate compute chart. Routing and shared components make “activated ÷ total” an incomplete performance model.
Million-token context
At a 1M-token context setting, the official model card reports that V4 Pro requires 27% of the single-token inference FLOPs and 10% of the KV cache used by DeepSeek V3.2. The comparison is publisher-reported under its implementation and setup; it is not a guarantee of your hosted provider’s latency, price, or memory.
A large maximum context also does not guarantee that every token receives equal attention or that an application accepts the full window. Retrieval quality, evidence location, prompt structure, reserved output, tool payloads, request caps, and operating cost all matter.
reported single-token inference FLOPs vs V3.2 at 1M
reported KV cache vs V3.2 at 1M
Always name the comparison baseline, context, metric, source, and date. Do not rewrite “10% of V3.2 KV cache” as “90% less memory everywhere.”
Training and reasoning
DeepSeek reports pretraining the two models on more than 32T diverse, high-quality tokens. Its post-training description uses domain-specific experts trained through supervised fine-tuning and reinforcement learning with GRPO, then unified through on-policy distillation. These are publisher descriptions; the public weight release does not by itself make the full data pipeline independently reproducible.
Fast, direct behavior for routine tasks. Lower explicit reasoning effort can reduce latency and output, but quality still needs task-specific validation.
More deliberate reasoning for coding, planning, math, and difficult analysis. Measure total tokens, latency, and accepted-result rate.
Highest published reasoning effort for boundary testing. The model card recommends at least a 384K context in local use; that is not a requirement for every ordinary prompt.
Benchmark evidence
The official model card separates Base, instruct, reasoning-mode, and frontier-model comparisons. Keep those tables separate. For example, an MMLU score for V4-Pro-Base is not a Think Max score, and a publisher-run SWE Verified result is not a guarantee for a different agent scaffold.
Open weights and source boundary
DeepSeek’s release calls V4 open-sourced, links an official weight collection, and the V4 Pro model card states an MIT license for the repository and model weights. Operators should still verify the license in the exact repository/revision, because rights attach to artifacts—not to a family name copied onto a third-party file.
The hosted API is a separate service with model IDs, pricing, limits, and terms. This website is another separate application layer. Publishing weights does not prove that a hosted endpoint uses the same revision or that a balance pack buys direct API credit.
The official model card describes both Pro and Flash as Mixture-of-Experts language models. It highlights a hybrid attention design using Compressed Sparse Attention and Heavily Compressed Attention, plus Manifold-Constrained Hyper-Connections and the Muon optimizer.
V4 Pro publishes 1.6 trillion total and 49 billion activated parameters. V4 Flash publishes 284 billion total and 13 billion activated. Total and activated counts answer different questions and should always be labeled.
In an MoE model, routing selects a subset of expert capacity for a token. This gives the model a large total weight set without using every expert for every token. It does not make cost, memory, or speed equal to a dense model with the activated count.
DeepSeek’s release, model card, and direct API table all publish a 1M context. A provider or application can impose a smaller request limit, and usable quality still depends on evidence placement, retrieval, latency, and memory.
DeepSeek calls the preview open-sourced and publishes official weights. The model card states the repository and weights use an MIT License. “Open weights,” API availability, and a fully reproducible training stack remain different claims, so inspect the exact repository and license.
The release links the DeepSeek V4 technical report hosted with the official V4 Pro repository. Use that report and model card for architecture and evaluation claims rather than an undated third-party summary.
No. DeepSeek-V4.io is an independent application. Model architecture facts describe the official V4 release; the website’s serving model is determined by its runtime provider configuration.
Source snapshot checked July 22, 2026. This is a technical reading guide, not an affiliation or independent reproduction of publisher experiments.
Compare the actual deployed workflow with representative prompts, tools, latency, usage, and a blinded quality rubric.
