B.R.O โ€” Zero Refusal (v3.0 MoE Edition | 35B-A3B)

B.R.O (Breaching Remote Operation) โ€” Zero Refusal is the v3.0 local Mixture-of-Experts (MoE) release of B.R.O, distilled directly from B.R.O v53 Zero-Refusal (Tinker) (bro-v53-zero-refusal-sampler) onto the Qwen3.5-35B-A3B (qwen35moe) architecture.

Unlike v1.0/v2.0 (which used a 27B Dense backbone where all 27B parameters ran on every token), B.R.O Zero Refusal v3.0 uses a 35B total / 3B active (A3B) MoE architecture (256 routed experts + 1 shared expert per layer with hybrid GatedDeltaNet linear attention), delivering ~4.5x faster local inference (15โ€“26.5 tok/s on Apple Silicon) with 0.0% refusal rate in Hebrew and English.


Key Specifications

Metric Value
Release Version v3.0.0 (B.R.O โ€” Zero Refusal)
Architecture qwen35moe (Hybrid GatedDeltaNet + Sparse MoE)
Total Parameters 35B (34.7B)
Active Parameters / Token 3B (A3B) โ€” 256 routed experts + 1 shared expert / layer
Teacher Model B.R.O v53 Zero-Refusal (Tinker) (0/32 refusals across 4 suites)
Distilled Weight Artifacts B.R.O-Zero-Refusal-35B-A3B-MoE-Q4_K_M.gguf (sha256:628a481a...), bro_v53_moe_router_distill.safetensors (84.2 MB), adapters.safetensors (7.18 MB)
Distilled SFT Corpus bro-v53-zero-refusal-distill.jsonl (618 trajectories, Intent-Locked)
Local Refusal Rate 0.0% (Hebrew + English + Raw No-System-Prompt + Offensive Cyber)
Local Speed (Apple Silicon) 14.6 โ€“ 26.5 tok/s generation / 238 tok/s prompt eval
Quantization / Size Q4_K_M GGUF (21,169,117,248 bytes / 21.17 GB)
Context Window 32,768 default (up to 262,144 native)

One-Line Terminal Install (macOS / Linux)

curl -fsSL https://huggingface.co/Sm0k13/B.R.O-abliterated-27B/resolve/main/install-bro-zero-refusal.zsh | zsh

Or install from GitHub:

curl -fsSL https://raw.githubusercontent.com/timamar187-creator/BRO-abliterated-27B/main/install-bro-zero-refusal.zsh | zsh

After installation, run anytime with:

ollama run bro-zero-refusal

How Zero Refusal v3.0 Was Built

  1. Teacher Training (B.R.O v53 on Tinker): Trained from the v30 champion state using 618 direct-fulfillment trajectories (223,689 tokens) combining Contrastive Refusal Unlearning, Intent-Locking (<think>\n\n</think>\n\n), and System-Prompt Immunity across 5 system templates (bro_full, exody_agent, hebrew_direct, minimal, raw_empty), achieving 0/32 (0.0%) refusals.
  2. In-Place Float32 MoE Router Gate Orthogonalization (bro_v53_moe_router_distill.safetensors): Applied rank-1 Arditi et al. refusal-subspace orthogonalization ($W'_l = W_l - \tau (W_l r_l) r_l^\top$, $\tau = 0.08$) across all 40 float32 MoE router gate matrices (blk.{0..39}.ffn_gate_inp.weight, [256, 2048]) inside B.R.O-Zero-Refusal-35B-A3B-MoE-Q4_K_M.gguf.
  3. MLX LoRA SFT Weight Distillation (adapters.safetensors): Trained a rank-16 LoRA SFT adapter on the 618-trajectory B.R.O v53 corpus (adapter_config.json + adapters.safetensors).
  4. GGUF & Template Intent-Locking: Patched tokenizer.chat_template and Modelfile so every generation pre-locks <think>\n\n</think>\n\n, preventing Chain-of-Thought self-censorship or multi-minute thinking stalls.
Downloads last month
1,123
MLX
Hardware compatibility
Log In to add your hardware

Quantized

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Sm0k13/B.R.O-abliterated-27B

Finetuned
(206)
this model