DeepSeek V4-Pro launches — 1.6 trillion parameters, 49 billion switched on
DeepSeek released V4-Pro (model name 0813) to its app, web interface and API on August 13, 2026, a mixture-of-experts model with 1.6 trillion total parameters of which 49 billion activate per token, a 1-million-token context window and a 384,000-token maximum output
The three lines
- V4-Pro-0813: 1.6tn total parameters, 49bn active per token, 1M-token context, 384K max output
- In preview since April, now generally available — deep-reasoning and instant-answer modes in one model
- An API price increase is scheduled for August 16, three days after launch
Key questions
- What is new in DeepSeek V4-Pro?
- The stated focus is agentic work — sustaining long coding tasks without human intervention. The published specification lists a 1-million-token context window, a maximum output of 384,000 tokens, tool calling and structured JSON output, and both a deep-reasoning mode and a fast-answer mode in a single model. The name 0813 encodes the release date, following DeepSeek's date-based versioning.
- What does 1.6 trillion parameters with 49 billion active mean?
- It is a mixture-of-experts (MoE) architecture. The model contains many expert sub-networks, and a router activates only a few of them for each token. Total weight is 1.6 trillion parameters, but roughly 49 billion — about 3.1% — participate in computing any given token. The effect is the knowledge capacity of a very large model at inference costs closer to a much smaller one, which is the structural reason DeepSeek has been able to price aggressively.
- Can it really handle a million tokens of input?
- It accepts them. Retrieving from them accurately is a separate question. Through 2026, benchmark work has consistently found a gap between advertised context length and the length across which frontier models reliably locate specific facts, with accuracy degrading substantially between the 200K and 1M ranges. This is an industry-wide pattern rather than a DeepSeek-specific one.
A model that entered preview in April became generally available on August 13. DeepSeek V4-Pro, model name 0813, shipped simultaneously to the app, the web interface and the API.
The most striking figure in the announcement was not a benchmark. It was the architecture.
1.6 trillion parameters in total. 49 billion switched on per token. That is 3.1%.
1. The published specification
| Item | V4-Pro-0813 |
|---|---|
| Total parameters | 1.6 trillion |
| Active parameters per token | 49 billion |
| Context window | 1,000,000 tokens |
| Maximum output | 384,000 tokens |
| Modes | Deep reasoning / instant answer, in one model |
| Features | Tool calling, structured JSON output |
| General availability | August 13, 2026 |
The sibling model rounds out the family. V4-Flash-0731, published on July 31, has open weights on Hugging Face and an API in public beta; it targets fast responses and agent-style workloads. V4-Pro takes the other end — hard reasoning, mathematics and code.
2. What "1.6 trillion, 49 billion active" buys
At first glance the ratio looks wasteful: 1.55 trillion parameters sitting idle on every token.
This is a mixture-of-experts design. The model holds many expert sub-networks, and a router selects a handful for each incoming token. The rest stay resident in memory without participating in the computation.
| Design | Knowledge capacity | Compute per token |
|---|---|---|
| Dense model | Proportional to parameter count | Proportional to parameter count |
| MoE model | Proportional to total parameters | Proportional to active parameters |
The trade delivers one thing: large-model knowledge at small-model inference cost. It is the structural basis for the price-performance position DeepSeek has held.
It is not free. All 1.6 trillion parameters must still be resident in memory, so serving requirements remain those of a very large model. This is emphatically not a model that runs on a laptop — the opposite direction from the recent wave of small models built for local execution.
3. This release is aimed at agents
What DeepSeek emphasised was not a leaderboard position but the ability to carry a long task without supervision. Three specification lines point that way:
- 1M-token context — an entire codebase plus the working history can sit in one request
- 384K-token maximum output — a single response can contain a large-scale code change
- Tool calling and JSON output — formats consumed by other programs, not by a human reader
Taken together, that is the specification of a model meant to be delegated to, not conversed with. Most major releases in the first half of 2026 moved the same way; V4-Pro is the Chinese entry in that column.
4. A price increase three days later
One condition is attached to this launch: an API price increase scheduled for August 16.
Three days between general availability and a repricing reads two ways. It could be ordinary introductory pricing to drive early adoption. It could also be a signal that serving a 1.6-trillion-parameter model is genuinely expensive. Distinguishing between them requires the size of the increase — and the size and scope of the increase could not be confirmed.
One secondary summary put V4-Pro's output pricing at roughly one-sixtieth of a leading frontier competitor. The original source and comparison date could not be verified, and in any case the comparison changes on August 16.
5. Where this sits in August
Listing only the model news this page has covered in August makes the cadence visible.
| Date | Release |
|---|---|
| August 6 | LG open-sources EXAONE 2.0, a 750B model |
| August 9 | OpenAI halts deployment of Astra over a first "critical" cyber rating |
| August 11 | Meta ships a 30B model that runs on a laptop |
| August 12 | Nvidia demonstrates voice AI with 0.48s interruption latency |
| August 13 | DeepSeek V4-Pro reaches general availability |
Five releases in eight days. July ran at the same rate.
6. What is unresolved
First, the weights. July's V4-Flash was published openly on Hugging Face. Whether V4-Pro follows could not be confirmed — and for a 1.6-trillion-parameter model that is the single highest-impact open question in this announcement.
Second, the benchmarks. The agent and coding figures are company-published. No third-party reproduction was available at the time of writing, and the gap between self-reported scores and field performance has repeatedly been an issue across the industry in 2026.
Third, that 1-million-token headline. Advertised context length and the length across which a model reliably retrieves are not the same number. Benchmark work through 2026 found accuracy falling substantially between the 200K and 1M ranges for nearly every frontier model. The distinction, and how it is measured, is set out in our companion piece on context windows.
Sources
- Qz — DeepSeek V4-Pro official launch (August 13, 2026)
- llm-stats — AI news: DeepSeek-V4-Pro-0813 released August 13, 2026
- Hugging Face — DeepSeek V4-Flash is now official: what changed in the 0731 build
- DeepSeekV4Pro — V4 GA reportedly delayed to mid-August 2026
- Morph — LLM context window comparison (2026)