2026-05-27 · 特辑|稀疏注意力价格战:中国开源模型如何把推理成本压到每百万 token 一美元 episode artwork

EPISODE · May 27, 2026 · 34 MIN

2026-05-27 · 特辑|稀疏注意力价格战:中国开源模型如何把推理成本压到每百万 token 一美元

from The Long View

三十分钟的技术入门特辑。从 Transformer 的注意力机制讲起,一步步走向本周技术圈最重要、但被主流财经媒体完全忽略的事件——DeepSeek V4-Pro 把 75% 降价永久化、小米 MiMo V2.5 降价 99%、MiniMax M3 预告 prefill 提速 9.7×、decoding 提速 15.6×。讲清楚什么是 KV 缓存、为什么它是推理成本的真正瓶颈、Compressed Sparse Attention 和 Heavily Compressed Attention 怎么把 KV 缓存压到上一代的 10%、滑动窗口加 SGLang HiCache 怎么把数据传输降到 1/7、MiniMax 的两阶段索引方案为什么先退回 full attention 又切回稀疏。然后讲架构性降价和促销性降价的本质区别、$1/M output token 这条线对 OpenAI/Anthropic/Google 意味着什么、为什么美国芯片出口管制反而把中国 AI 公司逼成了效率工程的世界冠军、稀疏注意力之后是什么(Mamba 混合架构、Cerebras/Groq/Rubin 专用硬件、1-bit/Ternary 压缩)。三个 takeaway。

Episode metadata supplied by the publisher feed · Published May 27, 2026

Embed this episode

Ready to play

2026-05-27 · 特辑|稀疏注意力价格战:中国开源模型如何把推理成本压到每百万 token 一美元

0:00 34:33

No transcript for this episode yet

We transcribe on demand. Request one and we'll notify you when it's ready — usually under 10 minutes.

No similar episodes found.

No similar podcasts found.

Frequently Asked Questions

How long is this episode of The Long View?

This episode is 34 minutes long.

When was this The Long View episode published?

This episode was published on May 27, 2026.

Can I download this The Long View episode?

Yes. Use the download control on the episode player to save the publisher-provided media file.
URL copied to clipboard!