Abacus.AI CEO Bindu Reddy claims 1.5 billion people in China use open-source AI, arguing the US underestimates the technology

Alignment externality is often overlooked here. When 1.5 billion people use models built on different safety frameworks, shaped by different cultural values and deployed without Western guardrails, you get a parallel AI ecosystem that diverges faster than policy can track. The US fixation on controlling the frontier misses that the frontier is already global and open.

@bindureddy X is *extremely* pro open source depending on what you're following. I think you just need to hit your algorithm with a stick

@bindureddy It's the hardware efficiency. Qwen3.6-27B hits 77.2% on SWE-bench Verified while running on a single consumer GPU. The local agent loop is getting ridiculously cheap.

@ECLresearch @bindureddy Which benchmarks? You're talking about deepmind, or the open source models?

@bindureddy The practical advantage of open models isn't ideology.

It's the ability to fine-tune, distill, route, self-host, inspect behavior, optimize cost, and deploy closer to the workflow.

For production AI, that flexibility can matter as much as raw benchmark performance.

@bindureddy If you want to see what those SWE-bench scores actually mean under the hood, this is a solid breakdown: https://leetllm.com/blog/swe-bench-deep-dive

@bindureddy @grok what does hallucinations mean

@bindureddy All USA corps have to do is have cheap tokens to get more training data. They got high centered on their quest for the AGI God and made enormous grotesque models that cost a fortune.

@KlepperCasey @bindureddy Probably not, 9 months behind in the AI race is a lifetime.

@bindureddy Curious how much of that Chinese open source is actually native tech versus just people wrapping llama weights.

@bindureddy We're betting on smaller, fine-tuned models for specific product use cases. Are you seeing China's open source strength translate more to domain-specific applications, or general-purpose foundation models?

@bindureddy ok

@bindureddy It is not about open source (x embraces this) and I doubt the honesty of China-based models.

@bindureddy Honestly the benchmark wars don't survive contact with 1.5B daily users. Real feedback at that scale compounds faster than any lab can iterate.

@bindureddy US AI scene is so busy measuring token throughput they forgot the actual goal. China's open source models are already good enough for 90% of use cases and getting cheaper by the week. Meanwhile Silicon Valley burns billions to shave 2% off a benchmark nobody asked about lol

For most use cases, there is absolutely no reason for anyone to use the frontier models like gpt 5.4 and above and opus 4.6 and above, because the likes of deepseek and qwen and even gemma gives you comparable results at zero to no cost and managable inference. This is now just about US vs China then anything else!

@bindureddy people sleep on iteration speed. open weights means you ship a fix today instead of waiting for the API to catch up. that matters a lot when you run actual products

Dear westerner , I am Chinese, thanks for the compliments , I personally I have tried both including closed source ones like ChatGPT, Claude, Gemini so on, and opened source ones qwen 3.6 27b and Gemma 4 26B ,the differences between closed source ones and opened sources is really significant, that’s why when I come to serious work I would use closed source ones, when I need to ‘play’ LLMs I would like enjoy those open sources LLMs

@bindureddy Ironically the US is embracing this monolithic centralized closed AI whereas China is embracing a decentralized, open source approach. Oh how the turns have tabled.

@bindureddy You are spot on about that shift. The gap in performance between top models is now tiny, while the price gap is massive.