ABOUT
An independent research initiative at an early stage.
rappidAI is an early-stage independent AI research initiative developing compact language-model experiments and documented training, evaluation and local-inference workflows.
WHAT RAPPIDAI IS
Learning through real model development.
rappidAI develops compact language-model experiments and investigates training pipelines, evaluation, efficient inference and practical local deployment.
PROJECT STRUCTURE
One initiative, one development project, one model series.
rappidAI is the research initiative. Lumen is the experimental development project for training and local-inference workflows. Quantum is the model series produced within that work.
WHAT RAPPIDAI IS NOT
Experimental work is not production capability.
rappidAI is not presented as a frontier-scale laboratory. Its pilot models are experimental, unreliable in factual output and not intended as production assistants or for high-stakes decisions.
- Pilot status
- Experimental
- Production use
- Not intended
PUBLISHED WORK
Two public experimental pilots.
rappidAI currently publishes two experimental German base-completion models: quantum-1-pilot and quantum-1.6-pilot. Both contain 49,295,872 parameters and are available as F16 GGUF files for local experimentation. The accompanying public repository documents the training, evaluation and GGUF-export workflow.
The public releases and source repository document the artifacts and intended workflow, but final run manifests and complete training logs are not linked. The releases do not demonstrate production readiness, reliable factual answering or competitiveness with larger general-purpose models.
WHY COMPACT MODELS
Capability should be considered alongside resources.
Local inference can reduce dependence on external services and may keep prompts on the user’s device. Actual privacy depends on the application, configuration and surrounding infrastructure.
Design target
Local and resource-efficient experimentation.
CURRENT DIRECTION
The Echelon model line.
The current public work covers Echelon architecture preflight, tokenizer validation and Garden data-pipeline smoke tests. Echelon Base and a later Echelon Chat stage belong to one model line; no trained Echelon model is public.
Explore the researchEXPERIMENTAL OUTPUTS
Limitations are communicated directly.
Research outputs may be limited, repetitive, factually unreliable and inconsistent. They should be inspected as experiments, not treated as production systems.
- Factual output
- May be unreliable
- Response quality
- May be inconsistent
- High-stakes decisions
- Not suitable
- Production assistant
- Not intended
EARLY-STAGE RESEARCH
Independent research, built from Berlin.
Jonas Désiré Cikemgil is a Berlin-based independent AI developer focused on compact language models, documented training pipelines, GGUF deployment and local inference.
Founder
Jonas Désiré Cikemgil
Founder & Independent AI Research Developer
- Compact language models
- Documented training pipelines
- GGUF deployment
- Local inference
Berlin, Germany