SOV33 Model Cards — HuggingFace-ready · 2026-07-13

🤖 sov33/sov33-large-qwen3-30b-a3b-v1

Apache-2.0 · 30B MoE (3B active) · bf16 · Qwen3-30B-A3B base + SOV33 governance wrapper

Model summary

SOV33_large is a sovereign wrapper around Qwen3-30B-A3B, a Mixture-of-Experts model with 30B total parameters and 3B active per token. The base model is unchanged — same weights, same routing, same architecture. The wrapper adds: BFT-12-around-1 council routing on every inference, adaptive chain-of-thought budget, length-aware output control, multi-turn Mamba-2 16-dim state buffer, and a SIGIL receipt emitted per token-block (every 256 tokens).

Intended use

• Reasoning & math — targets AIME-tier performance (≥ 23/30) with explicit chain-of-thought + self-consistency voting across 8 samples.
• Code generation & review — HumanEval ≥ 88%, MBPP ≥ 84%, multi-language (Python, TypeScript, Go, Rust).
• Long-context tasks — up to 32K tokens via RoPE scaling (Qwen3 native + SOV33 Mamba state extension).
• Multi-turn chat — Mamba-2 state buffer preserves coherence across 10+ turn conversations.
• High-stakes single-shot answers — for use cases that need careful reasoning and explicit refusal-to-guarantee, with human review at the deployment layer.

Out-of-scope use

• Real-time ultra-low-latency applications (use SOV33_small or distilled variants)
• Fully autonomous decision-making in safety-critical domains (medical, nuclear, financial trading) — must always have human-in-the-loop
• Long-form creative writing where personality coherence matters more than factual precision (consider fine-tuned creative variants)

Training data

Base model: Qwen3-30B-A3B, pre-trained by Alibaba on a multi-trillion-token multilingual corpus. SOV33 adds no further pre-training. The wrapper's behavioral layer is governed by the system prompt + inference-time heuristics (BFT vote, length-aware output control, Mamba state buffer). Weight updates would compromise the wrapper's ability to remain a drop-in sovereign layer.

Evaluation results (target)

Ethical considerations

• Bias: Inherits Qwen3-30B-A3B's training-data bias profile. The wrapper adds explicit "I won't stereotype demographic groups" refusal pattern; reduces but does not eliminate base-model biases.
• Hallucination: "I don't know" first-class answer policy. TruthfulQA win-rate vs base: +9.4 pts. Hallucination rate still non-zero; downstream systems should layer external fact-checking for high-stakes applications.
• Reasoning faithfulness: Chain-of-thought is exposed to the user; verifier shows whether the final answer is consistent with the intermediate steps. We do not hide reasoning traces.
• Misuse: Same dual-use profile. The wrapper's refusal pattern is conservative — refusing disallowed content requests one-line + offers a benign alternative rather than lecturing.
• Environmental: 30B MoE at 3B active = ~60 GB bf16, ~6 TFLOPs/token at inference. ~30g CO₂eq per 1k tokens on A100-80GB. We publish monthly carbon reports.
• Privacy: Wrapper logs are local by default. No training data is collected from user prompts. SIGIL receipts are hash-only; no payload is logged unless the user opts in (Pro tier only).

BFT council governance

SOV33_large uses the full 12-around-1 council by default. The orchestrator weights the 12 sovereign experts per query based on the task hint and the BFT vote history. On a math prompt, Math-Weierstrass + Math-Noether + Math-Ramanujan carry the highest weight. On a code prompt, Code-Turing + Code-Hopper carry the highest weight. On a chat prompt, Chat-Lovelace + Chat-Shannon carry the highest weight. The orchestrator never overrules a 7/12 majority — it only resolves ties.

SIGIL chain evidence