DeepSeek ships open-source DSpark to accelerate V4; mid-July launch and $7.4B raise

DeepSeek released DSpark, an open-source technology designed to dramatically accelerate deployment and inference of its forthcoming V4 models, signaling a strategic pivot toward becoming an efficient inference-infrastructure company rather than only a model lab. V4 uses a Mixture-of-Experts architecture and a 1-million-token context window and is billed as running intelligent models cheaply and at scale; alongside DSpark, DeepSeek is developing DeepSpec for further speedups.

V4 officially launches in mid-July with a notable pricing twist: time-of-day 'surge' pricing that roughly doubles rates during peak hours, plus ultra-low off-peak inference costs. The company also closed a $7.4 billion round at a $50B+ valuation, backed by CATL, Tencent, NetEase and JD.com — a heavyweight domestic Chinese investor syndicate that reinforces the US-China decoupling theme running through the week.

Developers zeroed in on licensing: V4's Apache 2.0 terms were called 'the material change' for commercial deployments, and its Chinese-SimpleQA score of 84.4 was cited as the first open-weight parity with closed models. There is also urgency around migration — legacy model strings deepseek-chat and deepseek-reasoner retire permanently July 24, prompting calls to update deprecated identifiers across LangGraph and n8n.

What to watch: whether DSpark's speedups hold up outside DeepSeek's own stack, real-world off-peak pricing versus Western APIs, and whether surge pricing pushes latency-sensitive workloads elsewhere.