🌍China's Open-Model Blitz Rattles US AI Labs
TL;DR
Chinese labs are undercutting US rivals on price and shipping open weights. DeepSeek's V4-Flash-0731 runs 284B params (13B active) in about 142GB of GPU memory and solves tasks for 3 cents versus GPT-5.6 Luna's 5, a 40% edge, while Alibaba and Moonshot pile on.
Chinese labs are undercutting US rivals on price and shipping open weights. DeepSeek's V4-Flash-0731 runs 284B params (13B active) in about 142GB of GPU memory and solves tasks for 3 cents versus GPT-5.6 Luna's 5, a 40% edge, while Alibaba and Moonshot pile on.
Key Points
DeepSeek V4-Flash-0731: 284B params, 13B active, MIT license
Fits in ~142GB GPU memory at FP4, runnable on one enterprise system
Costs 40% less than GPT-5.6 Luna: 3 cents vs 5 cents per solved task
DSpark speculative decoding baked directly into the weights
In mid-July, 6 of the top 10 ranked models came from Chinese firms
Why It Matters
Open Chinese models that match frontier quality at a third of the cost push US labs toward policy restrictions rather than price competition.
Quick Facts
Frequently Asked Questions
Why does this matter?
Open Chinese models that match frontier quality at a third of the cost push US labs toward policy restrictions rather than price competition.
What happened?
Chinese labs are undercutting US rivals on price and shipping open weights. DeepSeek's V4-Flash-0731 runs 284B params (13B active) in about 142GB of GPU memory and solves tasks for 3 cents versus GPT-5.6 Luna's 5, a 40% edge, while Alibaba and Moonshot pile on.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.