In January 2025, DeepSeek did what many thought impossible: it launched two open source models under MIT license that compete directly with the world's best proprietary models — at a fraction of the development cost.
The impact was immediate. Stock prices of chip and AI infrastructure companies plummeted. The premise that cutting-edge AI required billions in compute had been questioned.
DeepSeek V3 is the general-purpose model. With 671 billion total parameters in MoE (Mixture of Experts) architecture, it activates only 37 billion per token. It's a broad-purpose model — reasoning, text generation, code, analysis.
DeepSeek R1 uses the same architecture but specializes in step-by-step reasoning. It was trained with a technique called GRPO (Group Relative Policy Optimization), without relying on extensive human supervision — which drastically reduced training costs.
The benchmarks were decisive:
MATH-500 (advanced mathematical reasoning): 97.3% — DeepSeek R1
AIME 2025 (mathematics olympiad): 87.5% — DeepSeek R1-0528
Competitive coding: human-level performance on competition problems
DeepSeek R1-0528, an update launched in May 2025, brought significant gains in mathematics, logic, and coding. In some benchmarks, it surpassed models like GPT-4o.
Released in December 2025, V3.2 introduced something new: direct integration of reasoning into tool-use workflows. The model doesn't just reason — it reasons while using external tools.
V3.2-Speciale, a variant focused on mathematics, achieved gold medal level performance in three of the world's most difficult competitions in 2025: IMO (International Mathematics Olympiad), IOI (International Olympiad in Informatics), and ICPC World Finals.
Both models — V3 and R1 — are licensed under MIT, the most permissive license in the open source ecosystem. Unrestricted commercial use, modification, redistribution — all permitted.
This makes them especially attractive for companies that need to customize models without license restrictions.
For technology teams and datacenters, DeepSeek R1 represents a concrete shift: reasoning capacity equivalent to cutting-edge proprietary models, executable on own infrastructure, with controlled inference costs.
DeepSeek proved that training efficiency can substitute for raw scale. Better data and optimization techniques can accomplish more than simply increasing the number of parameters.
DeepSeek redefined what's possible in open source. R1 and V3.2 aren't cheap alternatives — they're direct competitors to the world's best models.
And with MIT license, they're available for anyone to build upon.
Sources:
