DeepSeek releases R2 reasoning model with open weights

DeepSeek released R2, the successor to its R1 reasoning model, with open weights under a licence permitting commercial use. R2 benchmarks significantly ahead of R1 on mathematical reasoning, code generation, and multi-step logical deduction tasks, while the open-weight release allows enterprises to run inference on their own infrastructure without routing data to DeepSeek's API endpoints. The combination of frontier reasoning performance with self-hosted deployment is unusual at this capability tier.

**What 'open weights' means in practice.** DeepSeek R2's weights are available for download, allowing deployment on hardware that the enterprise owns or controls. This means: no per-token API cost, no data egress to a third-party inference provider, and no dependency on DeepSeek's API availability. However, running a model of R2's scale (estimated 670B total parameters with a mixture-of-experts architecture, roughly 37B active per call) requires substantial GPU infrastructure — at minimum 4 H100s for reasonable throughput, making self-hosted deployment appropriate for large enterprises or specialist AI infrastructure providers rather than typical mid-market organisations.

**Relevance for APAC data residency constraints.** DeepSeek R2's open weights solve the data residency problem that has prevented many APAC enterprises from using cloud-hosted Chinese AI models. For organisations subject to PDPO, PDPA, APPI, or PIPL that want Chinese-language reasoning capability at frontier quality, a self-hosted R2 deployment on local infrastructure satisfies residency requirements that a DeepSeek API call would not. This is particularly relevant for Taiwan, Korea, Japan, and Singapore enterprises processing regulated financial or personal data in Chinese.

**Performance on APAC-relevant workloads.** R2's reasoning capability shows particularly strong results on structured legal and financial analysis tasks where multi-step inference is required — contract comparison, regulatory change impact assessment, financial statement analysis. These tasks are common in the APAC financial services and professional services sectors that AIMenta primarily serves.

**AIMenta's editorial read.** DeepSeek R2 is the most capable open-weight reasoning model available, and its availability changes the enterprise AI evaluation landscape for large organisations. For mid-market teams without H100 cluster access, the practical path is through a cloud provider that hosts R2 (Hugging Face, Together AI, AWS Bedrock) rather than self-deployment. Evaluate R2 for reasoning-heavy tasks where its benchmark advantage is most pronounced.