Keural-Lab-1031

An experimental checkpoint from Keural, based on Qwen3.8-Flash-Next.

Parameters	~65B total, ~6.7B active per token
Base	Qwen3.8-Flash-Next

Evaluation

Same harness for both models (lm-evaluation-harness with vLLM, full test sets, thinking enabled).

Benchmark	Keural-Lab-1031	Qwen3.6-35B-A3B
MMLU	0.798	0.839
MMLU-Pro	0.750	0.847
GPQA Diamond	0.678	0.803
GSM8K	0.930	0.963
IFEval	0.860	0.941

Usage

vllm serve mkd-minkeun/Keural-Lab-1031 --tensor-parallel-size 4 --max-model-len 32768

For llama.cpp, see Keural-Lab-1031-GGUF.

License

Released under the Qwen Community License 1.0, as a derivative of Qwen3.8-Flash-Next. Under clause 2, a company that runs a Model-as-a-Service or AI Work Assistant business needs a separate license from Qwen before commercial use.