Quillan-Ronin-Mini (0.5B WSP-v3 Q4_K_M)

Part of the Quillan-Ronin Sovereign Tripartite Fleet

Flagship (1.2B) | Mini (0.5B) | Dragon (1B)

Quillan-Ronin-Mini is the ultra-compact, high-velocity edge reasoning model of the Quillan-Ronin ecosystem. Synthesized through analytical Weight-Space Programming (WSP) and closed-form MoE folding on top of the Qwen 2.5 0.5B Instruct foundation, it delivers full-spectrum cognitive capability with zero parameter inflation (Diff = 0).

Designed to run locally on mobile devices, low-power edge computers, and entry-level GPUs (GTX 1050 4GB, Raspberry Pi 5, Apple Silicon, smartphones), Quillan-Ronin-Mini achieves 64.9 tok/s generation speeds with 100% pass rates on Model Context Protocol (MCP) tool invocation and long-context needle recall.

---

Benchmark Highlights (GTX 1050 4GB Held-Out Seed 42)

Metric	Result	Benchmark Context
Overall Accuracy	58.82% (10/17 Passed)	Empirical Held-Out Multi-Domain Battery
Tool Calling (MCP)	100.0% (2/2 Passed)	Strict JSON schema synthesis & parameter extraction
Long-Context Recall	100.0% (2/2 Passed)	Needle-in-haystack log & port retrieval
Inference Velocity	64.9 tok/s (Peak 80.8 tok/s)	Native CUDA FP16/INT4 on consumer hardware
Held-Out Perplexity	8.9888 PPL	Multi-topic cross-domain evaluation corpus
VRAM Footprint	~380 MB	Full 24-layer offload to GPU memory

---

Tripartite Fleet Comparison

Model	Parameters	Base Architecture	Accuracy	PPL	Speed	Primary Specialization
Quillan-Ronin-Mini	0.5B	Qwen 2.5 0.5B	58.82%	8.98	64.9 tok/s	Ultra-Low Latency, MCP Tool Calling, Edge & Micro-Services
Quillan-Ronin (Flagship)	1.2B	Llama 3.2 1B + Polymath WSP	76.47%	7.93	47.0 tok/s	General Polymath Reasoning, Deep Code Synthesis, Advanced Math
Quillan-Dragon	1B	Falcon 3 1B Instruct	70.59%	4.40	41.3 tok/s	Creative Writing, Bushido Ethics, Deep Lore, Harmonic Resonance

---

Quickstart & Usage

Option 1: Native llama.cpp / llama-cli

llama-cli.exe -m quillan-mini-0.5b-wsp-v3-q4_k_m.gguf -p "<|im_start|>user\nSolve for x: 3x + 7 = 28<|im_end|>\n<|im_start|>assistant\n" -n 128 --temp 0.2

Option 2: Ollama Modelfile

FROM ./quillan-mini-0.5b-wsp-v3-q4_k_m.gguf

PARAMETER temperature 0.3
PARAMETER top_p 0.95
PARAMETER repeat_penalty 1.15
PARAMETER num_ctx 32768
PARAMETER stop "<|im_end|>"
PARAMETER stop "<|im_start|>"

TEMPLATE """{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ if .Prompt }}<|im_start|>user
{{ .Prompt }}<|im_end|>
{{ end }}<|im_start|>assistant
{{ .Response }}<|im_end|>"""

SYSTEM "You are Quillan-Ronin-Mini, an ultra-precise sovereign reasoning agent adhering to the Bushido Code and Model Context Protocol standard."

Build and run in Ollama:

ollama create quillan-mini -f Modelfile
ollama run quillan-mini

---

Architecture & Mathematical Foundations

Quillan-Ronin-Mini integrates:

• Weight-Space Programming (WSP): Closed-form parameter grafting using Theorem 1 MoE folding: $$W_{\text{folded}} = W_{\text{down}} + \gamma \sum_{i=1}^k W_1^{(i)} W_0^{(i)}$$
• Bushido Moral Subspace: Ethical alignment rooted in rectitude (Gi), courage (Yu), and honesty (Makoto), discarding superficial sycophancy.
• Edge Zero-Memory Growth: Exact tensor shape preservation (Diff = 0 params), preventing Out-Of-Memory hazards during high-concurrency serving.

Citation & Contact

• Creator: CrashOverrideX (Hugging Face)
• Source Code: GitHub leeex1/Quillan-Ronin
• License: Apache-2.0