DeepSeek Drops V4 Pro: 1.6 Trillion Parameters, MIT License, Ready to Self-Host
DeepSeek's 1.6-trillion-parameter V4 Pro 0813 is generally available as MIT-licensed open weights, with large reported gains on agentic coding benchmarks over its preview.
On August 13, 2026, DeepSeek released V4 Pro 0813 — the general-availability version of its flagship model — and did something that still carries weight in the current AI landscape: it published the weights under the MIT license. The 66 safetensors shards are ungated on Hugging Face.
The model is a Mixture-of-Experts transformer with 1.6 trillion total parameters and 49 billion activated per inference pass. Its context window is one million tokens, and its high and maximum reasoning settings support outputs of up to 384,000 tokens. DeepSeek includes its DSpark speculative-decoding module in the release, an efficiency feature intended to accelerate inference by predicting likely continuation tokens.
From Preview to General Availability, and the Numbers That Changed
DeepSeek's journey to V4 Pro GA traces back to an April 2026 preview. The jump from preview to 0813 is not incremental.
On DeepSWE, an agentic coding benchmark that measures a model's ability to complete realistic software-engineering tasks end to end, DeepSeek reports that the preview model scored 12.8 and V4 Pro 0813 scored 62.7 — a gain of nearly 50 points. That is a material change in the vendor's reported results, though the numbers remain sensitive to the evaluation harness and settings.
DeepSeek reports 87.9 on Terminal Bench 2.1, which tests an agent's ability to operate in a real terminal environment. Its model-card table compares V4 Pro 0813 with Opus-4.8, Fable-5, Kimi K3, and GLM-5.2. On DeepSWE, V4 Pro 0813's reported 62.7 sits ahead of Opus-4.8 at 58.0 and GLM-5.2 at 46.2, and behind Fable-5 at 70.0 and Kimi K3 at 67.5.
The comparison is useful as a release signal, not a final leaderboard. DeepSeek evaluated public code-agent tasks with its own minimal Harness configuration using maximum reasoning effort, so teams should reproduce the workloads that matter to them before treating a benchmark as a procurement decision.
What "MIT License" and "Open Weights" Actually Mean at This Scale
The open-weight release is the part with the most durable consequences.
For regulated industries — healthcare, finance, and legal services — a standard concern about frontier LLMs has been sending sensitive data to a third-party API endpoint. Self-hosting a model of this caliber can change that calculus. A hospital system or financial institution can run a capable model on its own infrastructure, under its own data-governance rules, without an external API dependency.
The MIT license permits commercial use, modification, and distribution subject to its notice terms. Combined with the FP8 and FP4 mixed-precision deployment options documented by DeepSeek, the practical self-hosting barrier is lower than the raw download size initially suggests — though it is still a substantial infrastructure project.
For teams that prefer API access, V4 Pro 0813 is also available through providers. Open weights do not eliminate the operational work of serving a one-million-token model; they give teams a choice over where that work, data, and control reside.
A Pattern Worth Watching
DeepSeek V4 Pro 0813 is another example of a Chinese AI lab releasing open weights for a model intended to compete with leading closed offerings. The pattern is consistent enough to have stopped surprising people. Its implications have not fully landed.
The world's most capable AI models are no longer exclusively accessible through proprietary APIs. Anyone with enough compute can download V4 Pro 0813 and run it. That is a structural shift in where AI capability lives and who controls access to it — one with consequences for enterprise procurement, national AI strategy, and the model market that will take years to work through.
DeepSeek says a V4.1 is on the way. The pace is not slowing.
Sources
Hugging Face / DeepSeek — DeepSeek-V4-Pro-0813 model card: https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813
Hugging Face / DeepSeek — DeepSeek-V4-Pro architecture and evaluation materials: https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro
Reuters — DeepSeek releases official V4 Pro model as it steps up expansion: https://www.reuters.com/world/china/deepseek-releases-official-v4-pro-model-it-steps-up-expansion-2026-08-13/
Simon Willison — DeepSeek V4 Pro 0813: https://simonwillison.net/2026/Aug/12/deepseek-v4-pro-0813/
Digital Applied — DeepSeek V4 Pro GA release: https://www.digitalapplied.com/blog/deepseek-v4-pro-ga-official-release-2026
Ofox — DeepSeek V4 Pro 0813 pricing, weights, benchmarks, and API access: https://ofox.ai/blog/deepseek-v4-pro-0813-price-weights-benchmarks-api-access-2026/
MindStudio — DeepSeek V4 Pro benchmarks: https://www.mindstudio.ai/blog/deepseek-v4-pro-0813-benchmarks
Unite.AI — DeepSeek ships V4 Pro as its flagship model leaves preview: https://www.unite.ai/deepseek-ships-v4-pro-as-its-flagship-model-leaves-preview/