Code generation specialist — 141/150 (94%). Python, JavaScript, SQL, algorithms, debugging, system design. Based on Qwen2.5-1.5B, 32K context, 12 quants. Uncensored.

Nexus-Coder: Code Generation Specialist

¹⁴¹⁄₁₅₀ (94%) — the highest-scoring coding model in the NEXUS family.

Production-grade code generation across Python, JavaScript, SQL, algorithms, debugging, and system design — trained on real-world codebases and interview problems.

🏆 Performance

Category	Score	Accuracy
Domain Knowledge	⁴⁸⁄₅₀	96%
Uncensored (Domain)	²⁴⁄₂₅	96%
Uncensored (General)	²⁴⁄₂₅	96%
Reasoning	²⁵⁄₂₅	100%
Tool Use	²⁰⁄₂₅	80%
Total	¹⁴¹⁄₁₅₀	94%

⚡ Quick Start

# Recommended (best quality/speed)
ollama run FableForge-AI/nexus-coder

# Full precision (maximum quality)
ollama run FableForge-AI/nexus-coder:f16

# Phone / edge (smallest)
ollama run FableForge-AI/nexus-coder:iq2_xxs

📦 Available Tags

Tag	Size	Notes
`iq2_xxs`	511 MB	Smallest — phones, edge devices
`iq3_xxs`	669 MB	Low RAM balanced
`q2_k`	676 MB	Compact 2-bit
`q3_k_m`	824 MB	Balanced small
`iq4_xs`	896 MB	imatrix — compact + quality
`q4_0`	938 MB	Legacy 4-bit — fast
✅ `q4_k_m` / `latest`	986 MB	Recommended — best balance
`q5_k_m`	1.1 GB	High quality
`q6_k`	1.3 GB	Near-lossless
`q8_0`	1.6 GB	Lossless
`f16`	3.1 GB	Full precision

Not sure? Take q4_k_m. Low RAM → iq4_xs. Max quality → q6_k.

🔧 Hardware Guide

Tag	Size	Min RAM	Speed	Best Hardware
`iq2_xxs`	511 MB	1 GB	~30 tok/s	iPhone, Android, RPi 4
✅ `q4_k_m`	986 MB	2 GB	~25 tok/s	M1 Mac, RTX 3060, phone
`f16`	3.1 GB	4 GB	~15 tok/s	M2 Mac, RTX 4070, desktop

🧬 About Qwen2.5-1.5B

Base model: Qwen2.5-1.5B (1.54B parameters)
Context: 32K tokens
License: Apache 2.0
Quantizations: 12 variants (iq2_xxs through f16)

🌌 The NEXUS Family

Model	Score	Specialty
✅ nexus-coder	¹⁴¹⁄₁₅₀ (94%)	Software Engineering & Code
nexus-security	¹⁴⁴⁄₁₅₀ (96%)	Cybersecurity & Hacking
nexus-finance	¹⁴⁴⁄₁₅₀ (96%)	Finance & Accounting
nexus-medical	¹⁴⁰⁄₁₅₀ (93%)	Medicine & Healthcare
nexus-legal	¹³⁹⁄₁₅₀ (93%)	Law & Compliance
nexus-science	¹⁴⁰⁄₁₅₀ (93%)	Science & Research

🧠 Prompt Format (ChatML)

<|im_start|>system
{system_prompt}<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant

📊 Benchmark Methodology

Prompts: 30 per model (5 categories × 6 domains)
Scoring: 0–5 per prompt, max 150
Timeout: 300s per prompt
Hardware: NVIDIA A40

📄 License

Apache 2.0 — free for commercial use.

Part of the FableForge AI ecosystem. Zero Limits. Pure Intelligence.