DeepSeek, the Chinese artificial‑intelligence research lab, rolled out preview versions of its latest large language model family on Tuesday, introducing DeepSeek V4 Flash and DeepSeek V4 Pro. Both models employ a mixture‑of‑experts design that activates only a subset of parameters for each task, a strategy that trims inference costs without sacrificing capability.

Each model supports a massive 1 million‑token context window, a size that lets developers embed entire code repositories or lengthy documents within a single prompt. The smaller V4 Flash packs 284 billion total parameters, of which 13 billion are active at any moment. Its larger sibling, V4 Pro, boasts 1.6 trillion total parameters and 49 billion active ones, making it the biggest open‑weight model currently available. By comparison, Moonshot AI’s Kimi K 2.6 has 1.1 trillion parameters and MiniMax’s M1 sits at 456 billion.

DeepSeek says architectural tweaks give the V4 series a measurable edge over its predecessor, V3.2, which carried 671 billion parameters. The lab claims the new models have “almost closed the gap” with leading closed‑source systems on reasoning benchmarks. In head‑to‑head tests, V4‑Pro‑Max reportedly outperformed open‑source rivals across a suite of logical tasks and even edged out OpenAI’s GPT‑5.2 and Google’s Gemini 3.0 Pro on selected benchmarks. Coding competitions showed performance comparable to GPT‑5.4.

Knowledge‑based evaluations tell a slightly different story. The V4 models lag behind the latest frontier offerings—OpenAI’s GPT‑5.4 and Google’s Gemini 3.1 Pro—by an estimated three to six months in raw factual recall. DeepSeek acknowledges the gap, framing it as a “developmental trajectory” rather than a flaw.

Pricing is a standout feature. DeepSeek V4 Flash costs $0.14 per million input tokens and $0.28 per million output tokens, undercutting the cheapest tiers of GPT‑5.4 Nano, Gemini 3.1 Flash, and Claude Haiku 4.5. V4 Pro’s rates sit at $0.145 per million input tokens and $3.48 per million output tokens, still cheaper than Gemini 3.1 Pro, GPT‑5.5, and Claude Opus 4.7. Both models currently handle text only, lacking the multimodal capabilities—audio, video, image generation—found in many closed‑source competitors.

The launch arrives amid heightened scrutiny of Chinese AI firms. U.S. officials recently accused China of large‑scale intellectual‑property theft targeting American AI labs. DeepSeek itself has faced allegations from Anthropic and OpenAI of “distilling”—essentially copying—their models. Despite the controversy, the company positions the V4 series as a cost‑effective alternative for enterprises and developers seeking open‑weight solutions.

Dieser Artikel wurde mit Unterstützung von KI verfasst.
News Factory SEO hilft Ihnen, Nachrichteninhalte für Ihre Website zu automatisieren.