NewmodelNON-AI MPL-2.0
astra-guardian-depth
Astra: looped-transformer reference with recurrent depth and latent reasoning, in C11, Python and CUDA.
🧠 Collection
Models and Training
Language models, fine-tune checkpoints, reward models and the training infrastructure behind them.
NewmodelAGPL-3.0
sovereign-reduction-algebra
Array reduction algebra for an experimental SUBLEQ attention / bi-encoder, with exact J 9.7 and R gmp arithmetic.
NewmodelAGPL-3.0
tensor-roll
Recursive CUDA-Q model quantizer: compress a 30B-class teacher into an ~8B student under a hard parameter budget.
modelApache-2.0
hilbert-4b-baseline-GGUF
Hilbert 4B decision baseline as Q8_0 and Q4_K_M GGUF, with row-level decision agreement against BF16 (Q8_0: 99.3%).
modelmpl-2.0
lean-llm-starter
Starter kit connecting Lean 4 formal verification with LLM inference pipelines.
modelTri-License
nano-gemini
Minimal Gemini model implementation for local inference.
modelTri-License
pocketlearn
Compact learning framework for on-device model training and inference.
modelNVIDIA Community
snapkitty-merged
Nemotron Mini 4.2B GGUF (Q4_K_M). QLoRA fine-tune experiment on sovereign corpus. 2.6 GB.
modelAGPL-3.0
snapkitty-mlc
SnapKitty Machine Learning in C: a compact C autograd and ML core library.
modelNVIDIA Community
snapkitty-nemotron
Nemotron fine-tune checkpoint on sovereign domain data.
modelAGPL-3.0
snapkitty-nemotron-harness
Deterministic local harness for Nemotron. EmojiCode persona control, Prolog and Lean gates, sealed receipts.
modelBSL-1.1
snapkitty-transformer
Corpus classification and training gate with Datalog rules and x86-64 assembly. WORM-sealed receipts.
modelTri-License
sovereign-gemini-gguf
Zero-dependency GGUF binary parser and neural Graph IR for Gemini-class models without loading weights.
modelAGPL-3.0
sovereign-mamba
Mamba SSM architecture integration for sovereign inference with state-space model routing.
modelAGPL-3.0
sovereign-memory-twin
Recurrent Memory Twin Network: Evidence/Silence verdicts, WORM-sealed, with an IBM i ledger bridge.
modelTri-License
sovereign-mimo-4b
4B parameter code reward model pruned from MiMo-7B. FSM pipeline, ERE gates, WORM sealing. Runs local on RTX 3080.
modelTri-License
sovereign-qra
Deterministic routing tensor replacing softmax in MoE dispatch. Zero entropy, proved bijective. Lean 4 zero-sorry.
modelSovereign Source
sovereign-transformer
Transformer architecture for sovereign inference pipeline with deterministic execution.
modelTri-License
sovlm
Sovereign statistical language model. N-grams, Kneser-Ney smoothing, BM25 retrieval, QRNG sampling.
modelAGPL-3.0
summon
Pip-installable framework for sovereign fine-tuning. QLoRA, functional API, layered corpus builder, WORM chain sealing.
modelAGPL-3.0
symbolic-morphology
From-scratch symbolic learner for Latin verb morphology: hand-derived gradients, no ML framework, implemented and benchmarked across 9 runtimes down to NAND gates.
Commercial license
SnapKitty code is free and open under AGPL-3.0 for open-source use. Building a commercial product or service? A proprietary commercial license from Snapkitty Collective LLC lets you ship SnapKitty code without the AGPL's source-sharing and network-use obligations.