DeepSeek (V4)

What it is

DeepSeek is a Chinese AI lab whose V3 and R1 models made global headlines in January 2025 for delivering benchmark performance competitive with leading closed models, released as open weights, and reportedly trained at a small fraction of the compute cost frontier labs were assumed to require. Whatever the exact cost figures, the release was widely treated as evidence that the gap between open and closed frontier models had narrowed faster than most of the industry expected.

Where these numbers come from

DeepSeek entered public preview with V4 on April 24, 2026, in two variants: V4-Pro (1.6 trillion total parameters, 49 billion active per token) and V4-Flash (284 billion total parameters, 13 billion active per token), both with a 1-million-token native context window. The MMLU figure above reflects the prior V3/R1 generation — this entry should be updated via the admin dashboard once V4 reaches general availability with confirmed official benchmarks.

Notable traits

• Uses a large Mixture of Experts architecture with a very large total parameter count but a much smaller active parameter count per token.
• The R1 generation was specifically trained to produce visible step-by-step reasoning before its final answer, a technique associated with stronger performance on math and logic-heavy tasks.
• Fully open weights, which has consistently triggered rapid community adoption for self-hosting and fine-tuning within days of each release.