01
Efficient by design
Models should be evaluated not only by capability, but also by the resources they require.
INDEPENDENT AI RESEARCH
Smaller Models.Focused Intelligence.
rappidAI documents compact German-language model experiments, GGUF releases and local-inference workflows. The two public Quantum pilots are experimental base-completion models; Echelon is the current strategic model line and remains at pipeline and preflight stage.
The work is intended for developers, researchers, students and builders investigating small-model training, documented evaluation and local deployment with limited compute. The current pilots are research artifacts, not chat assistants or production systems.
Current focus: Echelon data-pipeline validation, reproducibility and transparent evidence boundaries.
Research index
Start with the question you want to answer.
The site separates model releases, research evidence, agent infrastructure and supporting documentation so each area can be understood on its own.
RESEARCH THESIS
We do not believe every useful AI system needs to be enormous.
Our work focuses on compact architectures, documented training workflows, efficient inference and transparent evaluation. The goal is not to imitate frontier-scale laboratories, but to investigate where smaller systems can be genuinely useful.
02
Local where possible
Local inference can make AI more accessible and controllable while remaining less dependent on network connectivity.
03
Open about limitations
Experimental models must communicate their weaknesses as clearly as their strengths.
FEATURED MODEL
quantum-1.6-pilot
Public experimental release
MODEL CARD
rappidAI · research
quantum-1.6-pilot
Parameter size
49.3M
parameters
49.3M released parameters against the 506M configured Echelon target.
Primary use
Research into documented continued-pretraining workflows
- Model type
- LlamaForCausalLM-style experimental completion model
- Context
- 512 tokens
- Vocabulary
- 16,384
A 49.3M-parameter experimental German completion model released after a reported continued-pretraining stage.
- Languages
- German-language experimentation
- Release
- Publicly available as an experimental F16 GGUF release; the public artifact is 98,990,560 bytes.
Experimental model. Not production-ready and not intended as a production assistant.
Published architecture
MODEL EVOLUTION
From 49.3M pilot releases to the quantum-1-echelon pipeline.
01
Public experimental release
quantum-1-pilot
Public 49.3M-parameter base-completion experiment released as an F16 GGUF.
02
Public experimental release
quantum-1.6-pilot
Public F16 GGUF release whose model card reports continued pretraining on 500M additional German tokens.
03
Pipeline and preflight stage — no trained model
quantum-1-echelon
Public base-architecture preflight, validated tokenizer and Garden pipeline smoke test; no trained model release.
CURRENT RESEARCH
Focused questions, practical constraints.
01
Training pipelines
Versioned architecture, tokenizer and data-preparation workflows with explicit evidence boundaries.
02
Local inference
Quantization, memory efficiency and deployment through lightweight inference runtimes.
03
Evaluation
Comparing base and continued-pretraining stages through fixed prompts, next-token metrics and documented limitations.
DOCUMENTATION STATUS
Evidence, gaps and reuse boundaries.
The resources hub turns “Documented clearly” into a public record of sources, reproducibility, data provenance, responsible use and unresolved information.
Published
Publications
Source-linked research notes that separate observations, negative results and open questions.
Partial evidence
Reproducibility
Pinned revisions, GGUF checksums, reference commands and explicit gaps in the public run record.
Partial evidence
Data & training
Dataset sources, configured targets, smoke evidence, filtering and missing final manifests.
OPEN RESEARCH
Open publication is part of the research process.
Where licensing and safety constraints allow, rappidAI publishes model artifacts, documentation, evaluation notes and implementation details so that work can be inspected and, when all required artifacts are available, reproduced.
FOUNDER
Independent research, built from Berlin.
Jonas Désiré Cikemgil is a Berlin-based independent AI developer focused on compact language models, documented training pipelines, agent infrastructure and local inference.
Jonas Désiré Cikemgil
Founder & Independent AI Research Developer
CONTACT
Interested in the research?
For technical discussions, collaboration enquiries or questions about the models, contact rappidAI directly.
cikemgil@rappidai-research.com