What this is
This week a post on LocalLLaMA (a Reddit community tracking open-source LLMs) hit the front page: user jazir55 noticed that over the past two weeks, the open-source world has only seen flagship model updates from A-tier players like Qwen, DeepSeek, and GLM (Zhipu), while Nvidia Nemotron (Jensen Huang's open-source line), Allen AI's Olmo, LongCat, and other "B-tier" players have collectively gone quiet.
So-called "B-tier" is open-source community slang. It refers to models with smaller parameter counts (typically 7B–32B), focused on specific scenarios or academic research, and not directly benchmarked against GPT-4 or Claude. They are cheaper, easier to deploy privately — and the tier that most enterprises actually use in production.
Industry view
Voices supporting the "divergence thesis": We have noticed that since H2 2024, open-source resources have clearly concentrated at the top. Every time the three Chinese vendors — Qwen, DeepSeek, and GLM — release a new model, it triggers viral-level discussion on GitHub, backed by Alibaba, DeepSeek, and Zhipu having enough compute and engineering teams to sustain continuous investment. Projects like Nemotron and Olmo depend on academic institutions or single-point teams, so their update cadence was never stable to begin with.
The counterargument also holds: Some practitioners see this as just a release cycle issue, not ecosystem shrinkage. Open-source models have traditionally followed a "save up for a big drop" pattern — six months of silence followed by a release that catches up with closed-source is the norm. Nemotron previously skipped multiple versions to scale up directly, and the Olmo team has explicitly said the next version will rebuild the data pipeline.
Our take: Both sides have a point, but the trend toward divergence is real. Chinese users in particular should pay attention — the density of Chinese majors at the open-source frontier has already surpassed their US counterparts.
Impact on regular people
For enterprise IT: B-tier models going dark means the choice set for private deployment is narrowing — either go with "large enough but expensive" options like Qwen or DeepSeek, or stick with older models.
For individual careers: Product managers and developers need to recalibrate their "open-source is usable" expectations for 2025. The probability of new models landing within a quarter is low.
For the consumer market: Slowing open-source progress may stabilize pricing for closed-source products like ChatGPT and Wenxin Yiyan, narrowing room for enterprise price cuts.