Qwen 3 and the Open Source Revolution

april 28th: alibaba released qwen 3.

this matters more than you might think.

what qwen 3 brings

a family of models from tiny (0.6B) to massive (235B parameters, MoE).

i work at a company that doesn't open source its frontier models (for safety reasons we can debate).

but i still appreciate what open source brings to the ecosystem:

1. accessibility not everyone can afford api costs. open weights let researchers and small teams experiment.

2. transparency you can actually see what's in the model. that matters for trust.

3. innovation open models get fine-tuned, modified, studied. knowledge compounds.

4. competition keeps everyone honest. you can't charge premium prices for mediocre models when alternatives exist.

deepseek, qwen, glm—chinese labs are open sourcing aggressively.

the motivation is partly different (domestic ecosystem building, international competition), but the effect is similar.

more capable models available to more people.

alibaba has been iterating steadily. qwen 1 → 1.5 → 2 → 2.5 → 3.

each generation is notably better. the improvement curve is impressive.

for practical use: qwen 3 models are good for coding, reasoning, and general tasks. worth trying.

i believe in safety considerations around frontier models. i also believe in open ecosystems.

these aren't necessarily contradictory. you can be thoughtful about what you release and still contribute to the commons.

the conversation about what to open-source and when is evolving. rightly so.

tested qwen 3-235b today. genuinely impressed. the options are multiplying.