DeepSeek Quietly Ships V4-Pro: 1.6 Trillion Parameters, 1M-Token Context, $0.87 Per Million Tokens
News Summary
DeepSeek has quietly shipped the finished version of its flagship large language model, DeepSeek-V4-Pro-0813, moving the system out of preview status with no blog post, no press event, and no formal announcement. The build appeared in the company's official API documentation and on Hugging Face, and it is now live for developers through the API and for consumers through chat.deepseek.com. As of 6:00 AM Eastern Time on August 13, 2026, the model is generally available, following an initial preview release of the DeepSeek-V4 family on April 24, 2026.
A Launch Without Fanfare
The rollout stands in sharp contrast to the moment that first put DeepSeek on the global map. When DeepSeek-R1 launched on January 20, 2025, it triggered a dramatic market reaction, contributing to roughly $593 billion wiped off Nvidia's market value within a week. The V4-Pro-0813 update carried none of that spectacle. Engineers and researchers monitoring the company's Hugging Face repository and API changelog were among the first to notice the new build had replaced the preview version, underscoring how routine large model updates have become for the fast-moving lab.
What Changed Under the Hood
DeepSeek-V4-Pro is built on a Mixture-of-Experts (MoE) architecture with 1.6 trillion total parameters, of which 49 billion are activated per token. Compared with its predecessor, DeepSeek-V3, which had 671 billion total parameters and 37 billion active parameters, V4-Pro represents roughly 2.4 times the total parameter count and about 1.3 times the active compute per token.
The most significant architectural upgrade is in long-context handling. The model's context window has expanded eightfold, from 128,000 tokens in V3 to 1 million tokens in V4-Pro, with a maximum output length of 384,000 tokens. This is made possible by a new hybrid attention mechanism that combines Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), alongside a technique the company calls Manifold-Constrained Hyper-Connections (mHC). According to DeepSeek's technical report, at the full 1-million-token context length, V4-Pro requires only about 27 percent of the single-token inference compute and 10 percent of the key-value cache memory needed by the earlier DeepSeek-V3.2 model.
The V4 family also includes a smaller sibling, DeepSeek-V4-Flash, a 284-billion-parameter model with 13 billion active parameters, designed for faster, higher-volume workloads. Both models retain the DeepSeekMoE framework and the multi-token prediction strategy introduced in V3, and both were released with open weights under the MIT License, along with a full technical report and Hugging Face model cards.
Pricing and Market Position
DeepSeek-V4-Pro-0813 is priced at approximately $0.435 per million input tokens and $0.87 per million output tokens through the API. That pricing is dramatically lower than comparable frontier models from Western labs, with some industry comparisons citing input and output costs for top-tier competing models in the range of $10 and $50 per million tokens respectively for input and output.
How It Stacks Up on Benchmarks
Independent replication of the vendor-reported figures has not yet been completed, so the following benchmark results should be treated as figures self-reported by DeepSeek. On SWE-bench Verified, a widely used coding benchmark, V4-Pro scored 80.6, putting it roughly on par with other leading models and only a fraction of a point behind top closed-source competitors. On SWE-bench Pro, a benchmark considered a proxy for production-grade coding capability, V4-Pro scored 55.4 percent, trailing the best closed-source models by close to nine percentage points. On Terminal-Bench 2.0, V4-Pro scored 67.9 percent, within about 1.5 points of the leading closed-source result.
Across a broader set of nine agent-oriented benchmarks where comparable scores exist, the strongest closed-source competitor led by an average of roughly 5 percentage points; when one outlier benchmark is excluded, that gap narrows to under 3 percentage points, with DeepSeek-V4-Pro actually leading on two of the nine measures.
Why the Quiet Rollout Matters
Industry observers note that the muted release style reflects a broader shift in how frequently frontier labs now ship model updates, and how routine such updates have become even when the underlying performance gains are substantial. For developers and enterprises evaluating large language models, the combination of near-frontier benchmark performance and a fraction of the per-token cost of leading Western alternatives positions DeepSeek-V4-Pro as a notable option for cost-sensitive, large-scale deployments, particularly those involving long-document analysis given the model's expanded 1-million-token context window.
What to Watch Next
Because the benchmark figures released alongside V4-Pro-0813 are vendor-reported, independent evaluators and third-party leaderboards are expected to publish their own replication results in the coming weeks, which will offer a clearer picture of how the model performs outside of DeepSeek's own testing environment. Developers interested in trying the model can access it immediately through the DeepSeek API or the Expert Mode and Instant Mode options on chat.deepseek.com, and researchers can review the full architecture details in the DeepSeek-V4 technical report published alongside the Hugging Face release.