Qwen3 on Dubesor Benchmark

So Qwen3-4B (thinking) beats the old much bigger Qwen2.5-32B-Instruct (non-thinking), Qwen2.5-14B-Instruct as well as Qwen3-8B (non-thinking). Qwen3-14B (non-thinking) model unsurprisingly beats Qwen3-4B (thinking), but also lands just below the older Qwen2-72B-Instruct. Qwen3-30B-A3B (non-thinking) tails R1-Distill-Qwen-32B (thinking only) which is pretty impressive since it should mean that it is able to deliver comparable quality without thinking, but more importantly Qwen3-30B-A3B (non-thinking) also beats the older Qwen2-72B-Instruct. QwQ-32B (thinking only) lands just above Qwen3-32B (non-thinking), but far below Qwen3-32B (thinking). Interestingly Qwen3-14B (thinking) and Qwen3-8B (thinking) both beat the old big Qwen2.5-Plus (non-thinking, API only) model. And finally, Qwen3-30B-A3B (thinking) tails the old biggest Qwen2.5-Max (non-thinking, API only) which is only beaten by Qwen3-32B (thinking) and the current biggest Qwen3-235B-A22B in both thinking and non-thinking modes. All in all, it looks as though the Qwen3-30B-A3B in non-thinking mode is a decent sweet spot somewhere in the middle and with thinking enabled it's a very competent contender, all with the higher inference speed thanks to MoE architecture as a bonus.