DeepSeek V4-Pro lands at 1.6T total / 49B active under MIT license — V4-Flash at 284B/13B active for cost-efficient frontier-tier reasoning
DeepSeek-V4-Pro (1.6T total parameters, 49B active) ships as the flagship model for maximum reasoning, coding, and agentic performance. V4-Flash (284B total, 13B active) ships as the cost-efficient option. Both are released under the MIT license — supporting commercial use, modification, and distribution with minimal restrictions. The frontier-tier open-source story is now MIT-licensed, not just permissive.
The MIT license choice is strategic. Through 2024-2025 frontier-tier open-source models used a patchwork of bespoke licenses — Llama's Community License, DeepSeek's prior license terms, Qwen's various restrictions. Each restricted commercial use in subtly different ways, generating real compliance friction for enterprises. DeepSeek V4 going MIT means an enterprise can deploy the model in any commercial context, modify it, redistribute modifications, and integrate it into proprietary products — no license review needed. That's the same license terms as React, PostgreSQL, and most of the foundational open-source stack the industry already runs on.
The MoE specifications — 49B active out of 1.6T total for Pro, 13B active out of 284B for Flash — matter for the deployment economics. Active-parameter count drives per-token inference cost; total-parameter count drives the model's representational capacity ceiling. V4-Pro's sparse activation pattern means it serves frontier-tier reasoning quality at roughly the inference cost of a 50B dense model. For self-hosting deployments, that's the difference between economically viable and prohibitively expensive. Expect DeepSeek V4-Pro to become the default open-source frontier choice for enterprise self-hosting through Q3.
Hugging Face Blog — Best Open-Source LLM Models 2026 → · O-Mega — Top 10 Open Source LLMs 2026 DeepSeek Revolution → · Stormap AI — Game-Changing Open Source AI Models of 2026 →