For context, Alibaba is behind Qwen 2.5, which is the series of LLMs for desktop/self-hosting use. Most of the series is Apache licensed, free to use, and they’re what Deepseek based their smaller distillations on. Thier 32B/72B models, especially finetunes of them, can run circles around cheaper OpenAI models you’d need an $100,000+ fire-breathing server to run… if OpenAI actually released anything for public use.
I have Qwen 2.5 Coder or another derivative loaded on my desktop pretty much every day. And because its open, the whole community contributes back to it.
So… Yeah, if I were Apple, I would’ve picked Qwen/Alibaba too. They’re the undisputed king of iDevice-sized LLMs at the moment, and do it for a fraction of the cost/energy usage of US companies. They’d have to be stupid to pick OpenAI or Anthropic, and even stupider to reinvent the wheel themselves (which they already tried. Their local “small” LLMs are kinda crap for their size).
For whatever reason, the Chinese companies tend to go with very permissive licensing, while Cohere, Mistral, Meta (Llama), and some others add really weird commercial restrictions (though this trend may be reversing). IBM Granite is actually Apache 2.0 and way more “open” and documented than even the Chinese tech companies, but unfortunately their models are “small” (3B) and not very cutting edge.
Another note: Huggingface Transformers (Also Apache 2.0, and from a US company) is the reference code to run open models, but there are many other implementations to choose from (exllama, mlc-llm, vllm, llama.cpp, internlm, lorax, Text Generation Inference, Apple MLX, just to name a few).
Some of the distillations are trained on top of Qwen 2.5.
And for some cases, FuseAI (a special merge of several thinking models), Qwen Coder, EVA-Gutenberg Qwen, or some other specialized models do a better job than Deepseek 32B in certain niches.
For context, Alibaba is behind Qwen 2.5, which is the series of LLMs for desktop/self-hosting use. Most of the series is Apache licensed, free to use, and they’re what Deepseek based their smaller distillations on. Thier 32B/72B models, especially finetunes of them, can run circles around cheaper OpenAI models you’d need an $100,000+ fire-breathing server to run… if OpenAI actually released anything for public use.
I have Qwen 2.5 Coder or another derivative loaded on my desktop pretty much every day. And because its open, the whole community contributes back to it.
So… Yeah, if I were Apple, I would’ve picked Qwen/Alibaba too. They’re the undisputed king of iDevice-sized LLMs at the moment, and do it for a fraction of the cost/energy usage of US companies. They’d have to be stupid to pick OpenAI or Anthropic, and even stupider to reinvent the wheel themselves (which they already tried. Their local “small” LLMs are kinda crap for their size).
How do you Apache license a LLM? Do they just treat the weights as code?
It’s software, so yeah, I suppose. See for yourself: https://huggingface.co/Qwen/QwQ-32B-Preview
Deepseek chose MIT: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
For whatever reason, the Chinese companies tend to go with very permissive licensing, while Cohere, Mistral, Meta (Llama), and some others add really weird commercial restrictions (though this trend may be reversing). IBM Granite is actually Apache 2.0 and way more “open” and documented than even the Chinese tech companies, but unfortunately their models are “small” (3B) and not very cutting edge.
Another note: Huggingface Transformers (Also Apache 2.0, and from a US company) is the reference code to run open models, but there are many other implementations to choose from (exllama, mlc-llm, vllm, llama.cpp, internlm, lorax, Text Generation Inference, Apple MLX, just to name a few).
Deepseek R1 is currently the selfhosting model to use
Some of the distillations are trained on top of Qwen 2.5.
And for some cases, FuseAI (a special merge of several thinking models), Qwen Coder, EVA-Gutenberg Qwen, or some other specialized models do a better job than Deepseek 32B in certain niches.