DeepSeek V4 and the February 17th Mega-Launch

February 17th, 2026, will be remembered as the busiest day in AI model history. Five separate companies launched major models within hours of each other, coinciding with the Lunar New Year holiday.

DeepSeek V4 is the headliner for the open-source community. The 1-trillion-parameter model introduces three architectural innovations: Manifold-Constrained Hyper-Connections for better gradient flow, Engram conditional memory for persistent context, and Sparse Attention for efficient processing of its 1M+ token context window. Early benchmarks show 98% on HumanEval and 96% on GSM8K, though independent verification is still pending.

At an estimated $0.10 per million input tokens, DeepSeek V4 would be the cheapest frontier-class model ever released — 50x cheaper than Claude Opus 4.6. The model is expected to be released with open weights under an Apache 2.0-style license, continuing DeepSeek's tradition of open releases.

The same day saw Anthropic ship Claude Sonnet 4.6, Alibaba release Qwen 3.5 (a 397B multimodal model supporting 201 languages), xAI launch Grok 4.20 Beta with a novel 4-agent collaboration architecture, and Cohere debut Tiny Aya — a 3.35B parameter model that supports 70+ languages and runs on smartphones.

The concentration of launches on a single date reflects both competitive pressure and strategic timing. For developers, the practical impact is a sudden expansion of options at every price point, from free open-source models to premium frontier offerings.