Sovereign Models and GGUFs
Collection
Small, local-first language models. Nemotron fine-tunes, reward models, deterministic routing, and GGUF exports. Everything runs on consumer GPUs. • 13 items • Updated
Sovereign MiMo-4B
4B parameter code reward model pruned from MiMo-7B-RL. Every score passes through ERE P1-P5 verification gates. Every decision is WORM-sealed.
Runs local on RTX 3080 (3.3 GB VRAM). No cloud. No vendor.
Overview
| Base Model | MiMo-7B-RL |
| Parameters | ~4B (pruned from 7B) |
| Architecture | 24L, 2048H, 16Q/4KV GQA, SwiGLU, RoPE |
| Quantization | GGUF Q4_K_M (~2.5 GB) |
| Max Context | 8192 tokens |
| License | AGPL-3.0 | BSL 1.1 | MIT |
How It Differs From Base MiMo-7B
- Pruned: 32L/4096H/32H -> 24L/2048H/16H (GQA 4KV)
- FSM Pipeline: 7-stage PREFLIGHT -> REASON -> SEAL -> RESPOND with HMAC-SHA256
- ERE Gates: P1-P5 verification on every score — halts on failure
- WORM Chain: Append-only SHA-256 audit on every decision
- CUDA Kernels: Custom RMSNorm, FlashAttn, SwiGLU, RoPE from hilbert
Usage
Ollama (recommended)
Python
From Transformers
Architecture
| Parameter | Value |
|---|---|
| Layers | 24 |
| Hidden | 2048 |
| Heads | 16Q / 4KV (GQA) |
| FFN | 5504 (SwiGLU) |
| Norm | RMSNorm |
| Position | RoPE (theta=10000) |
| Vocab | 32000 |
| Max Seq | 8192 |
| Reward Head | linear(2048, 1) -> sigmoid |
Verification
- WORM sealed: Every inference produces an append-only SHA-256 receipt
- ERE gates: P1-P5 verification on every score
- Lean 4: Routing tensor sovereign-qra proved zero-sorry
- Deterministic: Same input -> same output, always
Sovereign Stack Integration
| Component | Source | What it does |
|---|---|---|
| FSM | DEVFLOW-FINANCE | 7-stage pipeline with HMAC-SHA256 seals |
| ERE Gates | bert-agent | P1-P5 verification, halt on failure |
| WORM Chain | DEVFLOW-FINANCE | Append-only SHA-256 audit chain |
| CUDA Kernels | hilbert | Ampere-ready (RMSNorm, FlashAttn, SwiGLU, RoPE) |
| Routing | sovereign-qra | Deterministic dispatch, zero entropy |
| Training Data | sovereign-training-corpus | 882 WORM-sealed pairs |
Hardware Requirements
| BBQBADDIE Spec | |
|---|---|
| GPU | NVIDIA RTX 3080 (10 GB) |
| VRAM | ~3.3 GB (model + KV cache) |
| CPU | AMD Ryzen 7 7700X |
| RAM | 32 GB |
Intended Use
- Code quality scoring with verified reasoning chains
- Reward modeling for RLHF with deterministic audit trails
- Sovereign inference where every decision must be inspectable
Limitations
- Pre-training weights not yet available (architecture + pipeline ready)
- Benchmarks pending first training run
- Requires sovereign stack components for full ERE/WORM integration
Citation
Need exactly one file argument
License
Tri-license -- choose any one:
- AGPL-3.0 for open source / community use
- BSL 1.1 -> MIT for commercial / production use
- MIT after 2029-01-01
Copyright (C) 2026 Jessica L. Williams / SNAPKITTYWEST Bel Esprit D'Accord Irrevocable Trust