Model card · Public experimental release
quantum-1.6-pilot
49.3MParameter size
49,295,872 parameters
Published architecture
Summary
A 49.3M-parameter experimental German completion model released after a reported continued-pretraining stage.
- Status
- Public experimental release
- Model type
- LlamaForCausalLM-style experimental completion model
- Intended use
- Research into documented continued-pretraining workflows · Local German-language completion experiments
- Languages
- German
- Lineage
- The release card reports continued pretraining from Quantum 1 Base. The public configuration specifies weights-only initialization, a fresh optimizer, scheduler and step counter, and a frozen quantum-1 tokenizer; no final public run manifest verifies every configured detail.
- Release status
- Publicly available as an experimental F16 GGUF release; the public artifact is 98,990,560 bytes.
- License
- No model license is currently stated in the public repository. Downloadability does not by itself define reuse rights.
Technical dossier
Documented model facts.
Information linked to public project sources. Unknown release dates, licensing terms and unsupported client compatibility are not inferred.
- Version
- 1.6.0, matching the public quantum-1.6-pilot-v1.6.0-f16.gguf filename and release manifest
- Architecture
- LlamaForCausalLM-style; hidden size 512; intermediate size 1,536; 12 layers; 8 attention heads; 8 KV heads; tied embeddings
- Vocabulary
- 16,384 tokens
- Context
- 512 tokens
- Tokenizer
- Custom frozen quantum-1 tokenizer with a 16,384-token vocabulary
- Training data
- Approximately 100M German base-training tokens plus 500M additional German tokens, as documented in the public model card.
- Training configuration
- The public configuration targets approximately 30,518 steps, derived from 500M tokens at an effective 16,384 tokens per step. A final public run log is not linked.
- Evaluation
- The Hugging Face model card reports validation loss 3.348852 and perplexity 28.4700. No versioned evaluation report or standardized downstream-task benchmarks have been published.
- Release
- F16 GGUF, approximately 99 MB
- Quantization
- The public release provides an F16 GGUF; no Q8 or Q4 variants are listed
- Hardware
- Minimum RAM requirements have not been formally measured
- Inference
- llama.cpp compatibility is documented. Android comparison tooling exists, but no completed public Android validation report is currently linked.
Primary sources:Hugging Face model card ·Training documentation ·Generation diagnosis
Reference completion command
llama-completion -m quantum-1.6-pilot-v1.6.0-f16.gguf -p "Berlin ist" -n 64 --temp 0 --top-p 1 --top-k 0Completion prompting is the documented reference path. No chat template is promised.
Research context
How this release fits the experiment.
This release reports continued pretraining with a fixed architecture and tokenizer. Public configuration and code are available, but a final run manifest and complete training logs are not linked.
Repository reference
The pilot model card.
A visual snapshot of the model's parameter size and intended research use.
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
RELATED RESOURCES
Inspect the wider evidence.
Limitations
Read the limits first.
Semantically weak and factually unreliable output
Outputs may be incomplete or incoherent
Limited to a 512-token context window
No instruction tuning or chat alignment
No standardized task benchmarks have been published
Not suitable for medicine, law, finance, safety or other high-stakes applications
Not intended as a general production assistant
Third-party client compatibility has not been established by a published validation report
This model is experimental and not production-ready. It must not be used for high-stakes decisions.