DeepSeek DeepInfer follow-up: 155 trillion tokens a day on one cluster, KV cache kept for 7 days; researcher frames open vs. closed as 'light and darkness' (report)

Verified (fact)

InfoQ China covered DeepInfer, the inference framework DeepSeek released, and said in its headline that a single cluster can process about 155 trillion tokens a day and keep the KV cache for up to seven days. The South China Morning Post reported that DeepInfer supports running DeepSeek-V4.1-Flash on both NVIDIA SM100 (Blackwell) GPUs and Huawei chips. According to the same report, a senior DeepSeek researcher likened the gap between open-source and closed AI to a battle between 'light and darkness' and set a goal of 'total victory' for open source.

Needs fact-check (claim)

The figures of 155 trillion tokens processed per day and seven-day KV cache retention appear in a report headline; the measurement conditions have not been confirmed.

Why it matters (interpretation)

Production-scale throughput figures released alongside the code suggest the framework is a tool used in real operations, not research code. A KV cache kept for days also fits the push to lower the cost of long, multi-turn agent tasks.

Sources

• Press DeepSeek开源最新推理框架：单集群日吞吐155万亿Token，KVCache可保存7天 InfoQ China
• Press DeepSeek seeks ‘total victory’ over closed-source AI with new open-source tool South China Morning Post

Story thread

1. 2026-10-20 DeepSeek open-sources DeepInfer, the in-house inference framework behind its production services, with V4.1-Flash support (report)
2. 2026-10-21 DeepSeek DeepInfer follow-up: 155 trillion tokens a day on one cluster, KV cache kept for 7 days; researcher frames open vs. closed as 'light and darkness' (report)

More about DeepSeek

• 2026-10-23 P2 Policy Unconfirmed Safety & alignmentLanguage models DeepSeekMiniMax
  "DeepSeek and MiniMax complete registration under China's interim AI export rules, can keep open-sourcing abroad" (community report)

  A Reddit post, citing Chinese media reports, said DeepSeek and MiniMax have completed registration under the Chinese Ministry of...

• 2026-10-22 P2 Open source Unconfirmed KoreaLanguage modelsData & evaluation AlibabaGoogleMetaLG AI ResearchDeepSeekHugging Face
  Analysis: 49.3% of derivative language models posted to Hugging Face by Korean accounts this year are based on Chinese models, 46.8% on Qwen (community post)

  According to an analysis posted on GeekNews, a Korean tech news community, 49.3% of the 1,845 language models newly released on Hugging...

• 2026-10-22 P1 Other Unconfirmed Language models DeepSeek
  DeepSeek raises $15 billion days after US order to remove it from federal agencies (report)

  TechCrunch, citing reporting by another outlet, said DeepSeek has raised $15 billion, one of the largest investment rounds for a Chinese...

• 2026-10-21 P1 Infrastructure Unconfirmed Efficiency & infraLanguage models DeepSeekHuawei
  DeepSeek is using Huawei chips to pretrain its flagship model and is co-designing the chips (report)

  According to Korean media citing The Information, DeepSeek is working with Huawei from the design stage of its next-generation AI chips....

• 2026-10-20 P1 Open source Generally available Efficiency & infraLanguage models DeepSeekNVIDIAHuawei
  DeepSeek open-sources DeepInfer, the in-house inference framework behind its production services, with V4.1-Flash support (report)

  According to Korean media reports, DeepSeek on the 20th fully open-sourced DeepInfer, the inference framework that runs its large-scale...