Keural-Lab-1108

An experimental checkpoint from Keural, based on Qwen3.6-35B-A3B.

Parameters	~49B total, ~4.3B active per token
Base	Qwen3.6-35B-A3B

Evaluation

Same harness for both models (lm-evaluation-harness with vLLM, full test sets; GSM8K, IFEval and GPQA with thinking enabled).

Benchmark	Keural-Lab-1108	Qwen3.6-35B-A3B
MMLU	0.817	0.839
KMMLU	0.613	0.645
GSM8K	0.899	0.963
IFEval	0.916	0.941
GPQA Diamond	0.788	0.823

Usage

vllm serve mkd-minkeun/Keural-Lab-1108 --tensor-parallel-size 4 --max-model-len 32768

For llama.cpp, see Keural-Lab-1108-GGUF.

License

Released under the Apache License 2.0, as a derivative of Qwen3.6-35B-A3B (Apache 2.0).