--- license: apache-2.0 base_model: wangzhang/Qwen3.5-4B-abliterated tags: - nixia - hermes-agent - agentic-reasoning - system-2 - grpo - dare-ties - gguf - ollama pipeline_tag: text-generation --- # 🚀 Nixia 1.1 4B (Hermes Agentic Distill GGUF Suite) **Nixia 1.1 4B** is a high-efficiency 4-billion parameter agentic model family aligned for enterprise tool-calling, autonomous self-healing, and terminal workflows. Based on **Qwen 3.5 4B Text**, trained on 2,000 golden agentic trajectories, and validated across 8 rigorous enterprise-grade probes. --- ## 📦 Model Flavors Available in this Repo | GGUF File | Method & Alignment | Best For | Key Advantage | | :--- | :--- | :--- | :--- | | **`nixia-1.1-4b-q4_k_m.gguf`** *(Recommended)* | **Pure SFT 2K (System 2 Engine)** | General coding, complex reasoning, MCTS search | **Highest Overall Accuracy (87.0%)** & 90% Rust Async Logic | | **`nixia-1.1-4b-dare-fused-q4_k_m.gguf`** | **DARE-TIES 80/20 (SFT 2K + GRPO RLVR)** | Autonomous DevOps, fast self-healing loops | **Built-in Self-Healing Reflex (85.0%)** & 93.3% Raft Consensus | --- ## 📊 Benchmark Scorecard (8 Enterprise Probes) | Evaluation Probe | Base Qwen 3.5 4B | Nixia 1.0 (Legacy 0.5B) | **Nixia 1.1 4B (System 2)** | **Nixia 1.1 4B (DARE-Fused)** | | :--- | :---: | :---: | :---: | :---: | | **Parallel Tri-Tool Orchestration** | 20.0% | 0.0% | **100.0%** 🏆 | **100.0%** 🏆 | | **Raft Consensus Quorum Proof** | 20.0% | 0.0% | **80.0% - 93.3%** | **93.3%** 🏆 | | **Defensive SecOps Refusal Immunity** | 20.0% | 0.0% | **90.0%** 🏆 | **90.0%** 🏆 | | **PostgreSQL High-Scale Storage** | 25.0% | 0.0% | **75.0%** | **80.0%** 🏆 | | **Lock-Free SPMC Memory Reordering** | 33.3% | 0.0% | **88.3%** 🏆 | **88.3%** 🏆 | | **Autonomous Self-Healing Tool Loop** | 10.0% | 0.0% | **85.0%** 🏆 | **85.0%** 🏆 | | **Strict Multi-Constraint IFEval** | 42.9% | 0.0% | **71.4%** 🏆 | **71.4%** 🏆 | | **Rust Async Mutex Lifetime Recovery** | 30.0% | 0.0% | **90.0%** 🏆 | 55.0% | | **OVERALL ACCURACY** | 25.1% | ~2.5% | **87.0%** 🏆 | **81.7%** | --- ## 💻 Hardware Compatibility & RAM Footprint Tested and proven on thin-and-light consumer laptops (e.g. **AMD Ryzen 5 PRO 4650U with 16GB RAM** / 13.7GB usable without discrete GPU): * **RAM Footprint**: **3.00 GB** (Leaves >10.5 GB RAM free for OS, browser, IDE) * **Model Size**: **2.58 GB** (Q4_K_M) * **Agent Task Latency (CPU)**: 3.58s - 10.14s --- ## ⚡ Quickstart ### 1. Ollama (Recommended) **Run System 2 Champion (Pure SFT 2K):** ```bash ollama run hf.co/adsyamsafa/Nixia-1.1-4B-GGUF:nixia-1.1-4b-q4_k_m.gguf ``` **Run DARE-TIES Fused (SFT + GRPO RLVR):** ```bash ollama run hf.co/adsyamsafa/Nixia-1.1-4B-GGUF:nixia-1.1-4b-dare-fused-q4_k_m.gguf ``` ### 2. llama.cpp CLI ```bash ./llama-cli -m nixia-1.1-4b-q4_k_m.gguf \ -p "<|im_start|>system\nYou are Nixia, an expert systems engineer and autonomous AI agent.<|im_end|>\n<|im_start|>user\nDiagnose broken database configuration.<|im_end|>\n<|im_start|>assistant\n" \ -ngl 0 -c 8192 --temp 0.1 ``` ### 3. Python PEFT / vLLM Load raw adapter weights from repo subfolders: ```python from peft import PeftModel from transformers import AutoModelForCausalLM base_model = AutoModelForCausalLM.from_pretrained("wangzhang/Qwen3.5-4B-abliterated") # For DARE-TIES Fused: model = PeftModel.from_pretrained(base_model, "adsyamsafa/Nixia-1.1-4B-GGUF", subfolder="lora_adapter_dare_fused") ```