A Promising (?) Model Trained on Top of Qwen 3.5 397B

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10
1

Has anyone seen this? The benchmarks look impressive. GLM 5.1 performance for half the param count. 397B works well on 2 Sparks, so I think this should work too. The lab feels sketchy though. It seems to be some sort of AI focused private Chinese university. There’s only an MLX quant available for now.

They also have a Mini version that’s based on 35B A3B, with benchmarks matching the 27B. Not sure if I buy it.

2

Yep, looks good on benchmarks, fingers crossed that translates over to real use. So far I’ve not had much luck with finetunes, they always seem benchmaxxed and lose some smarts vs original weights.

3

Also, need someone to create the int4 autoround quantization. And port to vllm, since they recommend a modified sglang server.
Shame since 397b is a great model in dual sparks