Models developed through Hawking's research.

ModelSourceeBPWSizeTPSPrefillHumanEval+ARC-AGI-3Status
hwen3827-3Qwen3.8-27B (Qwen-27B)~311.02 GB (10.26 GiB)N/AN/AN/AN/AIn development

Metric definitions. eBPW is effective bits per weight. TPS is generated tokens per second. Prefill is prompt tokens processed per second. HumanEval+ and ARC-AGI-3 are public benchmarks. N/A means no qualified measurement of the exact artifact is available; it does not mean a score of zero.

Benchmarks and methodology

Results are published only for an exact model artifact, with hardware, batch size, context length and evaluation conditions stated. Coding results use EvalPlus HumanEval+; reasoning results use ARC-AGI-3. Hawking Systems' evaluation code and methodology notes will be published on GitHub with each qualified release.

Follow Hawking Systems on X and Hugging Face for future qualified releases.