Qwen 3.8 27b wins?
This is a Moe model, so, the performance isn’t better than the dense 27B qwen model.
Can you evaluate with general agentic, general knowledge, multilingual,… benchmarks. And compare to Qwen3.6 35B A3B?
Whether this model is a “Qwen3.8 35B A3B” model or not?
This is a Moe model, so, the performance isn’t better than the dense 27B qwen model.
Can you evaluate with general agentic, general knowledge, multilingual,… benchmarks. And compare to Qwen3.6 35B A3B?
Whether this model is a “Qwen3.8 35B A3B” model or not?
so you're saying that a dense model that is smaller is better because it's dense therefore shouldn't that scale a dense 1 Trillion would outperform all 1 trillion sparse MOEs?
A 27B dense has 27B forward activations each pass. The 35B-A3B MoE has 3B activations. For reasoning, as a rule of thumb it's fair to say an equivalent/closely sized will be smarter (and you pay for it with the 10X longer forward pass per token). There are some formulas people use for dense equivalency, but I don't take much stock in it, it's very much architecture/model dependent, eg from Qwen's own numbers, 3.5 27B beat the 3.5 122B-A10B in coding performance (SWE-bench verified, livecodebench).
An equally trained 1T dense would outperform a 1T sparse MoE, but no one is going to train (or run inference) on one.
BTW, I had some surprising results on my personal coding/agentic testing. Qwen 3.8 27B medium basically performed best - better than DS4 Flash 0731 (which I really like and trust). Ornith 1.5 is a step behind both and interesting not that much faster on my PRO 6000s, but it also seems to use a lot less tokens.
