Qwen3.8-Flash-Next: Time to Update Those Benchmarks
热度趋势
趋势数据积累中
百分比基于当前可用热度信号，而非评论数或独立用户人数。
Qwen 相关模型动态已经出现，适合跟踪能力变化、生态影响和后续可用性。
Qwen 3.8 27B 模型，特别是 Qwen3.8-Flash-Next 版本，在编码基准测试中表现出色，超越了大多数其他模型。然而，其通用知识能力相对较弱，不敌 Gemma 31B 和 Qwen 3.6。开发者正在积极收集更多与编码相关的数据以更新基准测试，因为模型日益复杂，这使得区分变得具有挑战性但令人兴奋。
it still very early, so had to disable oMLX K/V caching, qwen4_exp architectureis not yet supported + the obvious n-grams with which the whole 4 bit quant takes ~100G, so pretty tight
nevertheless, this is the first model for the year that was able to break through 94% on my cupel benchmark
one interesting bit is Qwen 3.8 27B is obviously great, but it did not do that well, since I have coding, general knowledge and science. it did outperform most in coding, but its general knowledge lost to Gemma 31B as well as to Qwen 3.6
this is the quant I tried with oMLX, which performed better than other 4 bit quants due to the mixed quantization:
this is a very good quant from Unsloth, it is not as strong as "MLX-mixed-4_8bit", but I could not fit a larger one from unsloth to be able to bench. You can see it on position #6 in the above leaderboard
I am working on collecting all I did for the last few months codingwise, and will add more pieces into the benchmark (hermes => pi / opencode, etc..) because models are getting too good to differentiate: I love it!