Chinese AI firm DeepSeek has quietly launched its latest large language model, V3, with no big announcement or white paper—just a simple upload to Hugging Face. The low-key release has nonetheless attracted widespread industry attention for one standout feature: it runs locally on Apple's consumer-grade hardware.
With 685 billion parameters—up slightly from its predecessor's 671 billion—V3 is widely seen as the base for the soon-to-launch DeepSeek-R2, an inference-optimized variant. DeepSeek typically rolls out foundation models first, followed by enhanced inference versions weeks later, and R2 is expected to follow that same playbook.
V3 brings high-end AI inference to consumer devices
V3 is released under the MIT license, allowing unrestricted commercial use. More notably, the model runs efficiently on consumer-grade machines, including Apple's Mac Studio with the M3 Ultra chip.
Apple machine learning researcher Awni Hannun confirmed that V3 can run at approximately 20 tokens per second on Macs equipped with the M3 Ultra chip, challenging the long-held belief that advanced AI models require enterprise-scale infrastructure.
High-end AI models have traditionally relied on Nvidia GPU clusters, often drawing thousands of watts. In contrast, the Mac Studio runs V3 at under 200 watts during inference, highlighting a significant shift in power efficiency and hardware requirements.
Developer and AI expert Simon Willison noted that V3's ability to run locally could signal a broader shift away from centralized data centers toward more decentralized, energy-efficient AI deployments across consumer devices.
Minimalist rollout, maximum disruption
DeepSeek also broke convention with its launch strategy—no white paper, no press coverage, just a silent model drop and an empty ReadMe on Hugging Face.
The no-frills release contrasts sharply with Silicon Valley's choreographed launches and reflects a pragmatic Chinese approach: prioritizing resource efficiency over media spectacle.
V3's hardware-friendly design could upend Wall Street's reliance on capital-intensive AI infrastructure, while its open-source nature aligns with China's broader push to close the gap with leaders like OpenAI, Anthropic, and Google DeepMind.
The release also lands amid intensifying competition. OpenAI plans to launch GPT-4.5 and an updated o3 inference model in 2025, while xAI's Grok, Anthropic's Claude, and Google's Gemini continue to evolve. Observers now speculate on how soon DeepSeek will deliver its next upgrade.
Reuters reports that DeepSeek's R2 inference model may debut ahead of its expected May timeline. Since R1 was trained on the V3 base, the recent release suggests R2's launch is imminent.
If R2 follows R1's path, it could position itself as a direct challenger to OpenAI's upcoming GPT-5, setting the stage for a high-stakes rivalry between open-source and closed-system approaches.
With US and Chinese AI capabilities converging, the global landscape is shifting. DeepSeek's open-source momentum mirrors Android's strategy of using scale and openness to challenge walled-garden ecosystems like Apple's.
It remains uncertain whether DeepSeek can overtake closed platforms—but its potential to disrupt the AI hierarchy is clear.
Article edited by Jack Wu
