Deepseek releases v4 open source models with one million token context

07:50

By: Dakir Madiha

Deepseek releases v4 open source models with one million token context

Chinese artificial intelligence company DeepSeek has unveiled preview versions of its V4 models, marking its most ambitious release since early 2025. The new models are open source and aim to rival leading proprietary systems at a significantly lower cost. The launch introduces two variants designed to handle large-scale tasks with extended context windows.

The flagship model, DeepSeek-V4-Pro, features 1.6 trillion parameters, with 49 billion activated per token. A lighter version, DeepSeek-V4-Flash, includes 284 billion parameters, with 13 billion active per token. Both models rely on a Mixture-of-Experts architecture and support a context window of up to one million tokens, placing them among the largest in the industry. A hybrid attention system reduces computational demands, allowing V4-Pro to use a fraction of the inference power and memory required by its predecessor.

Internal benchmarks show strong performance. V4-Pro scored 93.5 on LiveCodeBench and achieved a Codeforces rating of 3,206, surpassing models from Google and OpenAI in coding-related tests. On SWE-bench Verified, it reached 80.6 percent, close to leading models from Anthropic. Despite these results, the model still trails top proprietary systems in some knowledge and agent-based benchmarks.

Pricing is positioned as a key advantage. DeepSeek-V4-Flash costs $0.14 per million input tokens and $0.28 for output, while V4-Pro is priced at $1.74 and $3.48 respectively. This places it well below comparable models, making advanced AI capabilities more accessible to developers and enterprises.

The release follows delays linked to infrastructure challenges, including integration with domestic Chinese chips. The models are designed to run efficiently on Huawei Ascend hardware, supporting efforts to reduce reliance on foreign computing ecosystems. Major Chinese cloud providers such as Alibaba, ByteDance, and Tencent have reportedly prepared large-scale deployments. The models are now available via API and chat interface, with downloadable weights released under an MIT license.