SymbioticLM-14B
Model Type: Hybrid Symbolic–Transformer with Persistent Memory
Base Model: Qwen-14B
Framework: PyTorch + HuggingFace Transformers
Purpose: Full-scale cognitive reasoning model with self-organizing memory and generative symbolic evolution
Overview
SymbioticLM-14B is a 17.8-billion-parameter symbolic–transformer hybrid that couples high-capacity neural representation with structured symbolic cognition. It supports persistent memory, entropic recall, multi-stage symbolic routing, and self-organizing knowledge structures. This is an experimental research checkpoint — the capability claims below describe architectural intent, not benchmarked results (see Limitations).
This model is ideal for advanced reasoning agents, research assistants, and symbolic math/code generation systems.
Architecture Highlights
- Backbone: Qwen-14B transformer with rotary embeddings + FlashAttention
- Symbolic Dim: 8192
- Symbolic Modules:
- ThoughtDynamicsLNN (multi-head LSTM attention)
- LiquidThoughtProcessor
- CrystallineProcessor (DNAConv GNN)
- HelicalDNAProcessor (linear helical encoding)
- Memory: 4096 symbolic states in FP32, retrieved using entropy + contextual similarity
- Dream Mode: Background symbolic simulation for open-ended cognition
- Router: Intent classifier + entropy gating for processor path selection
Files Included
| File | Description |
|---|---|
model.bin | Transformer weights (LFS) |
model.safetensors | Memory-safe weights, optimized for loading |
memory.pt | 4096-symbolic vector bank |
config.json | Model and architectural metadata |
generation_config.json | Top-p, temperature, decoding settings |
tokenizer.json | Full tokenizer with symbolic tag support |
added_tokens.json | Tags like <D_LIM>, <PROOF>, <BY_MEASURE>, etc. |
special_tokens_map.json | Special token mapping for tokenizer |
Intended Uses
- Multi-step conversational agents with true memory
- Long-form symbolic theorem generation and proof planning
- Scientific dialogue, symbolic simulations, math/code synthesis
- Reasoning in fuzzy, discontinuous, or non-smooth problem domains
Limitations
- Memory requires curation and seeding for maximum utility
- Symbolic cognition is not instruction-tuned for general QA
- FlashAttention and symbolic modules increase VRAM usage during generation
Citations
Please cite "SymbioticLM" when using symbolic memory components in research or applications.
Convergent Intelligence Portfolio
Part of the Symbiotic AI Series by Convergent Intelligence LLC: Research Division
Mathematical Foundations: Discrepancy Calculus (DISC)
SymbioticLM's persistent memory and symbolic evolution connect to Discrepancy Calculus through self-generating completeness (Ch. 3 of the DISC monograph) and symbolic-root domains. The discrepancy operator:
$$Df(x) = \lim_{\varepsilon \downarrow 0} \frac{1}{\varepsilon} \int_x^{x+\varepsilon} \frac{|f(t) - f(x)|}{|t - x|}, dt$$
quantifies local mismatch between integration and differentiation. In the symbolic-transformer context, $D$ measures the gap between what the symbolic system encodes (discrete structure) and what the transformer integrates (continuous representation). The self-generating completeness theorem establishes that completeness emerges dynamically via energy computation on symbolic-root domains — the mathematical foundation for why symbolic-neural hybrids can produce structure that neither component generates alone.
The discrepancy energy $E_{\text{disc}}[f] = \frac{1}{2}\int w(x)(Df(x))^2 d\mu(x)$ provides a natural stability criterion for the memory consolidation process: memory states with bounded discrepancy energy are stable; those with divergent energy indicate structural transitions requiring reorganization.
Full theory: "On the Formal Analysis of Discrepancy Calculus" (CIx, 2026; Convergent Intelligence LLC: Research Division).
Related Models
Part of the SymbioticAI series — symbolic–transformer hybrids with persistent memory: Symbiotic-1B · Symbiotic-8B · Symbiotic-Beta. See the SymbioticAI collection for the full family.
From the Convergent Intelligence Portfolio
DistilQwen Collection — Our only BF16 series. Proof-weighted distillation from Qwen3-30B-A3B → 1.7B and 0.6B on H100. Three teacher variants (Instruct, Thinking, Coder), nine models. The rest of the portfolio proves structure beats scale on CPU. This collection shows what happens when you give the methodology real hardware.
Top model: Qwen3-1.7B-Coder-Distilled-SFT
Full methodology: Structure Over Scale (DOI: 10.57967/hf/8165)
Convergent Intelligence LLC: Research Division