Quin Beta 21.ab1 (8B) - [INTERNAL RESEARCH ARTIFACT]
Model Identifier: Quin-Beta-21.ab1
Status: EXPERIMENTAL (v0.2.1) - Unstable. Intended for architectural validation and internal testing of optimization techniques.
Research Context & Abstract
Quin Beta 21.ab1 (8B) is a pre-production, internal research artifact, not a general-purpose foundational model. It has been developed and released by Dr. Young and Seo Park Jun exclusively to serve as a public testbed for novel architectural hypotheses and optimization strategies.
This 8B sub-model is designed for the preliminary validation of new technologies, specifically the viability of Resonant Gating Units (RGUs) and Causal State Tensors (CSTs), before their costly integration into the scaled Quin-91 (910B) parent framework.
This release serves as a public-facing checkpoint for internal ablation studies. It is built from scratch to provide a clean baseline for evaluating these new techniques, isolated from the complexities of a scaled-up, fully-trained model. This model is intended for personal research and academic review of the core architecture only.
RGU Activation (Conceptual)
The RGU activation function $\phi(x)$ is governed by a resonance function $R$ which modulates the input $x$ against a learned frequency matrix $\Omega$ and a gate $G$. This forces the network to find "resonant" pathways for data flow.
The conceptual formula for a single RGU block is:
$$\phi(x) = G(x, W_g) \cdot \tanh(R(x, \Omega) \odot (x W_v))$$
Where the resonance $R$ is calculated as:
$$R(x, \Omega) = \cos^2(\text{proj}(x, W_k) \cdot \Omega)$$
This non-linear, frequency-dependent gating is the core mechanism being tested in the Quin-21 beta.
Model Description
Quin Beta 21.ab1 is an 8-billion parameter, decoder-only model built on the experimental Quin-Architecture. It is a high-fidelity, parameter-projected sub-model derived from the primary Quin-91 (910B) research trunk.
The 21.ab1 (Asymmetric Beta 1) revision was built from scratch and completed a preliminary training cycle of 9 days. Its sole purpose is to validate the stability and inferential scaling of the core architectural components.
The Quin architecture diverges from standard transformers by replacing the feed-forward sub-layer with a Resonant Gating Unit (RGU) and utilizing Causal State Tensors (CST) for context management.
Architectural Innovations (Quin-Arch v0.2b)
The core hypothesis of the Quin-Arch is that parameter efficiency can be logarithmically scaled by treating the model's latent space as a predictive vector field rather than a simple information store.
- Asymmetric Parameter Projection (APP): The 8B parameters are not a sparse subset but a high-fidelity projection mapped from the 910B parent's latent state space. This mapping is designed to allow
Quin-21 to inherit conceptual properties and dimensional weightings that far exceed its raw parameter count.
- Resonant Gating Units (RGU): RGUs replace the standard FFN/MoE sub-layer. They utilize a dynamic, input-derived frequency-matching algorithm to selectively activate parameter "pathways." This resonance-based activation significantly reduces inferential compute (FLOPs) per token by activating only the most salient neurons, while maintaining high representational density.
- Causal State Tensors (CST): This model does not use a conventional Key-Value cache. Instead, it generates a Causal State Tensor ($T_{cs}$) for each token. $T_{cs}$ is a 4th-order tensor encapsulating a predictive vector field of potential future states, allowing the model to "anticipate" context rather than just "remember" it.
Model Details
| Parameter | Specification |
|---|
| Model Type | quin (Custom Architecture) |
| Variant | 21.ab1 (Asymmetric Beta 1) |
| Parameter Count | ~8.1B (Projected from 910B) |
| Parent Model | Quin-91 (910B) |
| Context Window | 8192 Tokens (Truncated) |
| Gating Mechanism | Resonant Gating Unit (RGU) |
| State Management | Causal State Tensor ($T_{cs}$) |
Development Status & Disclaimers
- PRIMARY OBJECTIVE: This model was developed for personal research by Dr. Young and Seo Park Jun to test novel optimization strategies and architectural designs in a controlled, low-parameter environment.
- NOT FOR PRODUCTION USE: The model is unstable, non-optimized for inference, and may produce unpredictable or non-coherent outputs. It has not undergone any alignment, instruction tuning, or safety training.
- INTERNAL TOOL:
Quin-21.ab1 functions as an internal testbed. Public availability is provided for architectural transparency and academic discussion only. DO NOT ATTEMPT TO USE THIS MODEL FOR ANY REAL-WORLD APPLICATION.
Intended Use (Theoretical)
As this model is an ab1 beta artifact intended for internal R&D, direct transformers integration is not supported. Loading requires a custom QuinLoader class (not publicly available) that can correctly initialize the CST and RGU layers.
Hypothetical Loader Implementation (Illustrative Pseudocode):
1# WARNING: Pseudocode. This loader does not exist in the public library.
2# This code is for conceptual demonstration of the architecture's requirements.
3
4from quin_arch_loader import QuinLoader, QuinConfig
5
6# Configuration must specify RGU expansion and CST dimensions
7config = QuinConfig(
8 model_type="quin",
9 variant="21.ab1",
10 n_layers=32,
11 rgu_expansion_factor=2.75,
12 cst_dimensions=1024,
13 projection_parent_hash="quin-91/proj-map-v2.bin"
14)
15
16# The loader would build the model from the custom config
17model = QuinLoader.from_pretrained(
18 "drchamyoung/Quin-Beta-21.ab1",
19 config=config,
20 device_map="auto",
21 precision="bf16"
22)
23
24# print(model.get_architecture_report())
25# > [Quin-21.ab1]: 8.1B Params. 32 RGU layers. CST dim 1024.
