Expired50% confidenceFactExact time
Currently, more than 70% of DeepSeek's computing power is used for training, with inference accounting for less than 30%.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
9/25/2026
First Seen
This claim has expired. It is kept for the record and no longer describes the current state.
Valid until: 10/9/2026(expired)
Sources
Related Entities
Related Claims
UnverifiedDeepSeek将后训练阶段的强化学习计算量提升到了预训练的10%左右，而大多数公司这一比例仅为1%-5%75% similarUnverifiedModels specifically trained on code (such as the DeepSeek-Coder series) often score 10–30 percentage points higher on code benchmarks than general-purpose models with the same parameter count.72% similarUnverifiedDeepSeek open-sourced DSpark (repository name: DeepSpecs), which achieves 50% to 400% inference speedup without retraining or quantization.71% similarUnverifiedDeepSeek's innovations in Multi-head Latent Attention (MLA) and FP8 mixed-precision training compress KV Cache memory usage to 5%–13% of standard multi-head attention.71% similarUnverifiedDeepSeek-V3以极低的训练成本实现了顶尖推理能力70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/948673API
curl https://kongchang.com/api/v1/knowledge/claims/948673MCP
get_claim(id=948673)