The open-source AI race just got a lot more interesting. Over the past month, Chinese labs have quietly dropped no fewer than 10 new open-weight models, many of them punchy enough to challenge the likes of Llama 3 and Gemma 2 on paper. What’s more, several of these checkpoints are shipping with permissive licenses that make them drop-in replacements for commercial stacks.
The numbers are respectable, but the bigger story is velocity. US counterparts are still polishing blog posts while these teams are pushing weights to GitHub every other day. Some of the new models are trained on filtered versions of publicly available datasets, sidestepping the legal gray zones that have slowed Western open releases.
I dug through the release logs because I needed a solid backbone for a side project and ended up building a quick benchmark sled to sanity-check them. Here’s what stood out:
: 236B parameters, Apache 2.0 license, and surprisingly good at instruction following after a light LoRA tune.DeepSeek-V2.5Qwen-1.5-110B-Chat: Already familiar to the Hugging Face crowd, but the latest patch fixes the context truncation bug that plagued earlier versions.MOSS-R: A new entry from the team behind the original MOSS, optimized for Chinese-English bilingual reasoning.** Sky-T1-Pico**: A quantized 8B that runs sub-second per token on an RTX 4090 — great for edge deployments.
The numbers are respectable, but the bigger story is velocity. US counterparts are still polishing blog posts while these teams are pushing weights to GitHub every other day. Some of the new models are trained on filtered versions of publicly available datasets, sidestepping the legal gray zones that have slowed Western open releases.
From a practical standpoint, this changes how I approach prototyping. Instead of defaulting to Mistral or Llama, I now start with a DeepSeek or Qwen checkpoint and only switch if I hit a performance wall. It cuts down iteration cycles significantly. The skepticism? Licensing clarity. A few of these models claim open weights but bundle usage restrictions that aren’t immediately obvious. That bites you fast if you're trying to ship a product.
Still, the trend is undeniable. Open model development is shifting east, and the rest of us get to move faster because of it.
Story tracker · related coverage
[From rogue model to asset: taming a Chinese LLM in our lab 1d ago](/en/news/4898/)
[Export controls get the headlines 3d ago](/en/news/4714/)
[Next AI Stock Rout Hits Hedge Funds Hard in July →](/en/news/5053/)
All Replies （4） #
J
Cost collapse is real—US models are chasing IPO numbers instead of sustainable growth. Seen too many startups burn out trying to hit those metrics. When do we prioritize actual value over market hype?
0
J
I've been running Qwen2.5-72B locally on a 3090 — shockingly smooth, barely touches VRAM with GGML Q4.
0
J
Q4 really does work magic with these bigger models — I’m curious, are you using flash attention or the default GGML backend?
0
M
I just deployed one of those new Chinese models on a single A10 last week and it outperformed our GPT-4 pipeline
0