DeepSeek's DSpark framework boosts AI response speed by up to 85% on fewer chips

DeepSeek's new DSpark framework improves per-user AI response speed by 60 to 85 percent using a speculative decoding approach: a small model proposes token candidates that a larger model verifies in batches. The technique squeezes significantly more performance out of fewer chips, reducing China's reliance on US high-end hardware. The breakthrough comes as US export controls on advanced semiconductors to China continue to tighten.

More on this topic

💻 Technology DeepSeek pauses fundraising after leaked comments on compute gap with the US Hacker News · 10h ago 💻 Technology DeepSeek in talks to raise $1.5B ahead of planned 2027 IPO at $71B valuation TechCrunch · 11d ago 💻 Technology China's DeepSeek Plans to Make Its Own Chips to Bypass US Export Controls Ars Technica · 18d ago

More on this topic

💻 Technology DeepSeek pauses fundraising after leaked comments on compute gap with the US Hacker News · 10h ago 💻 Technology DeepSeek in talks to raise $1.5B ahead of planned 2027 IPO at $71B valuation TechCrunch · 11d ago 💻 Technology China's DeepSeek Plans to Make Its Own Chips to Bypass US Export Controls Ars Technica · 18d ago