Research
Open questions, bounded claims.
Research at Cinqic stays legible: implementation, measurement, historical results, and limitations are not the same thing.
01 — Research record
Evidence before the headline.
Research at Cinqic is documented through repositories, validation results, and final conclusions. Published work keeps implementation, measurement, uncertainty, and limits distinct.
02 — Released model
Juniper 1 Base
Juniper 1 Base is Cinqic’s first original base language model, trained from scratch. It is a 142.77-million-parameter dense decoder-only Transformer: 20 layers, width 768, SwiGLU feed-forward blocks, grouped-query attention with 12 query and 6 key-value heads, rotary position embeddings, and tied embeddings, with a 24,576-token Cinqic tokenizer. It was trained for one epoch on the j1-base-v0.2.2 corpus (about 2.95 million training documents, 3.41 billion training targets) at a 2,048-token context. The released checkpoint is the frozen juniper1-base-v1, training step 52,034, selected under a recorded protocol. Its full validation-split loss is 2.43 nats per token (0.916 bits per byte); that figure was observed on the validation split during training and is not an independent capability benchmark.
Juniper 1 Base is a next-token base model only. It is not instruction-tuned and makes no chat, instruction-following, tool-use, or reasoning claims. Post-training experiments on it (supervised fine-tuning versions 2, 2.1, and 2.2, and one distillation attempt) did not pass their registered gates; they are recorded as closed, failed results and are not part of the release. Juniper 1 research and development is closed with this release. The weights, tokenizer, training code, and the frozen selection record are published under the Apache License 2.0; dataset materials are published to the extent the source terms permit.
Separate from AAA
Juniper 1 Base was trained in its own repositories, independently of AAA. It is not integrated into AAA, no AAA research result depends on it or applies to it, and Juniper app’s model support is a separate product capability.
Read the release repository ↗Model card on Hugging Face ↗03 — Current research
AAA — Accurate Autonomous Adaptation
AAA is Cinqic’s experimental research program on systems that learn from feedback, adapt when conditions change, keep what remains useful, and recognize when their knowledge is no longer adequate. Its tests are built so that the answer can be no.
Python programming is the current specialization. The current phase, aaa.python.v1, is evidence-gated work before any scaling: on fresh held-out tasks, a structural code representation beat a lexical one at the same size, and capacity helped up to about 4,000 trainable parameters. A roughly 10,000-parameter version did not earn its extra cost, so scaling is not justified yet. Simple deterministic methods still match or beat every learner on some tasks.
Scoped world-model and language experiments followed. A world model improved decisions within one generated debugging benchmark, with documented seed and transfer limits. A small language-model result is recorded as not verified, because an independent review found that test feedback crossed the language channel’s boundary. The earlier moving-dot benchmarks are kept as historical evidence and regression tests, including their failures. A proposed next AAA phase, an Erudition Model that would decide when and how a system’s components adapt, exists only as a research pull request under separate-session review; it is not merged, and no result from it is claimed here.
Claim boundary
AAA makes no claim to general intelligence, understanding, or independent goal formation. Model size is not treated as progress; capacity is added only when the evidence justifies it. AAA is separate from Juniper 1 Base: that model is not integrated into AAA, and no AAA result applies to it.
Read the AAA repository ↗04 — Retired research
Juniper LM 1
Juniper LM 1 was a controlled study of modernizing the original GPT-2 124M checkpoint while keeping its architecture, tokenizer, and size fixed. Development ended on October 4, 2026 (America/New_York).
The stock GPT-2 baseline, frozen benchmark versions, failed experiments, research logs, and provenance remain preserved as historical evidence. Retirement does not invalidate JuniperBench-Code v1.2 provenance.
Historical record
No Juniper LM 1 checkpoint was trained or released. The research record remains available as a concise historical summary.
05 — Retired research
Juniper Encoder
Juniper Encoder was a small encoder-only model investigated for routing requests among registered agents, tools, and skills. It was retired on September 15, 2026 with research complete: the configured benchmark matrix was executed in full, but held-out routing accuracy did not support deployment.
Historical record
No model was released and no production checkpoint was selected. The repository is public and read-only, retained for research history and reproducibility.
Read the retirement record ↗06 — Completed research
Juniper Math 1
Juniper Math 1 is a completed and retired research project. Its final conclusion found that the tested small architecture and training regimes learned meaningful component behaviors, but no Phase 8 checkpoint simultaneously met the project’s direct capability, tool-use, retention, and reliability gates.
Historical record
It is not an active model project, and no Phase 8 production or release checkpoint was approved.
Read the final conclusion ↗