ConeML builds, adapts, licenses, deploys, and maintains language models for organizations. Its public R&D program makes the underlying model-engineering and evaluation methods inspectable.
Commercial work can include:
Engagements begin with the workflow, available material, deployment constraints, unacceptable failures, and acceptance test—not with an assumption that new training is required. Public research-model licenses do not grant commercial rights; customer work, ownership, licensing, and support are scoped by separate agreement.
ConeML's public releases demonstrate the engineering and evaluation methods behind its commercial work. They publish measured strengths, failure boundaries, shipped-artifact checks, peer context, and resource assumptions. They are evidence of method, not claims of customer deployment.
The fourth public ConeML model introduces a scratch-trained 810M family with an 8,192-token context. ConeML 810M Alpha favors executed single-function writing and two-step tasks; ConeML 810M Alpha-Arithmetic records a stronger four-lane arithmetic profile. Both ship as Hugging Face weights plus F16 and tested Q8_0 GGUFs.
The model cards publish task-family-matched strengths beside weak GSM8K and HumanEval results, peer wins, resource assumptions, and known limitations. Public research models are released for non-commercial use under CC BY-NC 4.0 unless stated otherwise.