DeepSeek V4 Pro 0813: Architecture, API Pricing, and Deployment Limits
DeepSeek's V4 Pro report describes a 1.6T-parameter MoE model with 49B active parameters. Current API pricing varies by peak and off-peak periods; see the rate card and deployment caveats.

DeepSeek's V4 Pro report describes a 1.6T-parameter MoE model with 49B active parameters. Current API pricing varies by peak and off-peak periods; see the rate card and deployment caveats.
Written and checked by the RankLLMs editorial team. See theeditorial policy.
DeepSeek V4 Pro’s published technical report describes a 1.6-trillion-parameter Mixture-of-Experts model with 49 billion activated parameters per token. The distinction matters: active parameters relate to computation for a token, while storing and serving the full model still requires memory for the full set of weights and runtime state.
This overview separates research-report architecture from API pricing and deployment. Pricing below reflects DeepSeek’s current API pricing page as checked on October 9, 2026. The page labels the Pro model version DeepSeek-V4-Pro-0813; rates vary by peak/off-peak period. DeepSeek’s September 10 V4.1-Flash announcement also described a temporary routing of the deepseek-v4-pro API name to V4.1-Flash beginning September 14 until V4.1-Pro launched. Because documentation and routing can change, confirm the model identifier and version returned for your account before relying on a price or capability.
Architecture in context
| Item | Reported specification | What it does and does not mean |
|---|---|---|
| Total parameters | 1.6T | Approximate scale of the full MoE weight set |
| Activated parameters | 49B per token | Portion selected for computation; not the total weight storage requirement |
| Context | Up to 1M tokens | Maximum documented context; practical performance varies by task and serving setup |
| FLOPs comparison | 27% of V3.2 in a stated comparison | A report-specific comparison, not a universal efficiency ratio against all models |
| KV cache comparison | 10% of V3.2 in a stated comparison | Applies to the report’s stated configuration and comparison baseline |
The report discusses compressed-attention and residual-path methods designed to reduce inference and long-context costs. A percentage such as “10% KV cache” is meaningful only with its baseline and setup; it should not be generalized to every model, sequence length, or deployment.
Current API rate card
DeepSeek prices API tokens by model, cache hit or miss, and time window. Its current pricing page lists:
| Token type | Off-peak | Peak |
|---|---|---|
| Input, cache miss (per 1M) | $0.66 | $1.32 |
| Input, cache hit (per 1M) | $0.022 | $0.044 |
| Output (per 1M) | $1.98 | $3.96 |
The page defines off-peak as half of peak pricing and gives the applicable UTC schedule. It says prices may change. Budget with the rate shown at time of use and account for the input/output mix, cache behavior, and any provider-specific routing. The older $0.435 input and $0.87 output figures previously shown here are not the current rate card.
Reading benchmark results
Benchmarks can help compare models only when the dataset, prompt, tools, scaffolding, sampling, and scoring version line up. A reported Terminal-Bench score is evidence about a particular test setup, not a promise of performance on every repository or agent framework. Review the original evaluation methodology before treating any single score as a deployment forecast.
Deployment and memory
The 49B active parameter count does not mean the model fits in 49B worth of VRAM. Every parameter still has to be stored somewhere across devices, even if only a routed subset is used per token. As a simple lower-bound illustration, 1.6 trillion one-byte weights occupy about 1.6 TB before quantization details, metadata, runtime buffers, or KV cache. At FP8, practical memory needs exceed the raw weight calculation. A workstation-class pair of GPUs cannot hold the full unquantized model; serving requires an appropriate multi-device setup, offload strategy, or hosted API.
For teams that cannot support the full deployment, the API is the practical route. Confirm the returned model name and published rate, then test the workload on representative tasks. For self-hosting options at smaller scales, consult our open-source model guide.
Related resources
Takeaway
V4 Pro’s report describes a large sparse model and techniques intended to improve long-context efficiency. Those architectural gains do not eliminate the memory needed to store its full weights. Its current API price is also time-dependent, and a recent routing announcement makes it especially important to verify the model actually served. Use the live docs and your own representative evaluation before projecting price or quality.
Frequently Asked Questions
How many parameters does DeepSeek V4 Pro have?
DeepSeek's V4 technical report describes Pro as a 1.6-trillion-parameter MoE with 49 billion activated parameters per token. Active parameter count is not the total memory required to store model weights.
What does DeepSeek V4 Pro cost through the API?
The current pricing page lists Pro input at $0.66 per million cache-miss tokens and $0.022 per million cache-hit tokens off-peak, and $1.32/$0.044 at peak. Output costs $1.98 off-peak or $3.96 at peak per million tokens. Rates may change; check the live page.
How should I compare DeepSeek V4 Pro benchmark results?
Check the model snapshot, dataset version, agent harness, tools, and scoring procedure for each result. Scores from different setups are not directly comparable, and a benchmark score does not guarantee performance on your workload.
Can I run DeepSeek V4 Pro on a workstation?
The full model has 1.6 trillion parameters, so storing all weights requires substantial aggregate memory. At a rough one byte per parameter, the weights alone are about 1.6 TB before runtime overhead; a small workstation is not enough for the full model.
Source links for data and claims discussed in this article are listed below. Refer to each source for its current figures and methodology.
- •DeepSeek-V4 technical report(source)
- •DeepSeek API models and pricing(source)
- •DeepSeek V4 Pro GA release(source)
- •DeepSeek V4 model collection(source)
Recommended Reading
Related by topic and recency.
DeepSeek V4 Flash 0731: Current API Alias and Historical Context
DeepSeek's API now serves legacy V4 Flash aliases as V4.1-Flash. See the current context limits and peak/off-peak prices, and why old snapshot benchmarks need separate evidence.
Grok 4.6: Capabilities, Benchmarks, and Current API Pricing
Grok 4.6 is a 500K-context xAI model for coding and agentic work. See xAI's current short- and long-context pricing and what published benchmark comparisons can show.
Best LLMs for Coding in 2026: How to Compare Them
Choose a coding LLM by matching the benchmark to your workflow, checking the agent harness, and measuring task quality and cost on your own repository.
