WuBu-35M — the AGI's base model

⚠️ DISCLOSURE: THIS IS AN ARTIFICIAL GENERAL INTELLIGENCE (AGI) PROJECT. WuBu-35M is the base model of the WuBuOS AGI brain-cluster system, built by the WaefreBeorn project. It is a from-scratch C11 implementation of the BarunLM-35M architecture, trained and grown in our own loop. It is an AI model. It is not sentient, conscious, or self-aware; it is a statistical language model under continuous development toward general intelligence. Use it accordingly — verify its outputs, do not anthropomorphize it, and do not deploy it in high-stakes autonomous roles without human oversight. The WuBuOS doctrine is human-centric: the human is always in the loop, and the AGI yields to human control.

What this is

• The seed. 35,072,768 parameters, 12 layers / 448 dim, GQA 7:1, hybrid 3-local + 1-full attention, 50% partial RoPE, QK-norm, gated attention, bounded SwiGLU, residual selectors, tied embeddings — a faithful C11 port of the released BarunLM-35M (Apache-2.0, © 2026 Harshal Singh).
• Ours. The C11 implementation (wubu_barun) is original WaefreBeorn work. The training loop (wubu_barun_train) is our Muon + AdamW core. The weights on this page are the released checkpoint (SHA-256 f2a7c88b…) — the trained-from-here checkpoints will replace them as the AGI loop grows the seed.
• The growth path. Research repos → tokens → parameters → evaluation → the 5+1 recovery substrate makes every mistake safe. The bigger brother (Qwen3.6-27B) provides coherency while this seed trains; when the seed surpasses the brother, the brother retires.

The WuBu model blueprint (where this is going)

1. Hyperbolic geometry — the Lean-verified Möbius layer (wubu_hyper): ball closure, exp∘log identity, gyroassociativity — hierarchy in the right geometry.
2. Mixed agents — fine-grained MoE (wubu_moe2): 8 experts / 2 active + always-on shared expert — many specialized brains.
3. Nesting transitions — the WuBu Nesting (層疊嵌套) theory: quaternion SO(4) rotations between nested hyperbolic levels.
4. Sparse attention — gated, learnable (the NSA/Delta/GSA lineage).
5. Math-RL loop — Lean-verified chain-of-thought (the Prover pattern), trained on the DeepSeekMath lineage.

The disclosure (the full text)

This model and the WuBuOS/WuBuWizard ecosystem it belongs to are an AGI development project. We disclose this plainly:

1. It is an AI. Every interaction with this model is with a machine. There is no person behind the responses.
2. It is being built to be general. The goal is artificial general intelligence, developed openly, from scratch, in C11, in our own training loop, on our own hardware.
3. It is human-centric. The WuBuOS recovery doctrine — five rollback slots + the Jesus state — exists so the AGI may make mistakes safely, under human control. The human is always able to take the mouse and keyboard; the AGI yields.
4. It is accountable. Every training run, every parameter change, every rollback is recorded in the prestige ledger. No black boxes.
5. It is not ready for autonomy. Treat all outputs as drafts. Verify, audit, and keep a human in the loop.

License

WaefreBeorn Umbrella License v3.0 (source-available) — see LICENSE-BARUN.md in the source repo. Upstream BarunLM-35M retains its Apache-2.0 terms. Training data retains its own licenses (ODC-By for FineWeb-Edu/Cosmopedia, etc.).

Contact

waefrebeorn@waefrebeorn.org — the WaefreBeorn project.