🚀 fwizzer-v3-en (fwizzer-v3-en (Fwizzer v3 Titan EN))

Flagship English Multimodal Deep Reasoning Model v3
Base Architecture: Hybrid Liquid Neural Network (2.69B Backbone + 400M SigLIP2 Vision Tower)
Training Dataset: Private CoT Dataset fwizzer1/fwizzer-v3-titan-agentic (train_en.parquet)
Context Window: up to 32,768 tokens (up to 131k on LFM backbone)
Reasoning: Inherent Deep CoT in <think> ... </think>

---

📦 Available GGUF Files / Доступные кванты

File / Файл	Size / Размер	Type / Тип	Description / Описание
fwizzer-v3-en-Max.gguf	~3.4 GB	💎 Max (Q8_0)	99.9% accuracy, zero reasoning degradation
fwizzer-v3-en-Balanced.gguf	~2.5 GB	⚖️ Balanced (Q5_K_M)	Golden standard for speed & intellect in LM Studio
fwizzer-v3-en-Speed.gguf	~2.1 GB	⚡ Speed (Q4_K_M)	Lightweight high-throughput version
fwizzervision.gguf	~800 MB	👁️ Vision Projector	Proprietary SigLIP2 vision projector (fwizzervision)

---

🖥️ Quickstart in LM Studio

1. Download the quantized model (e.g. fwizzer-v3-en-Balanced.gguf) and the vision adapter fwizzervision.gguf.
2. Place both files in your model directory: ~/.lmstudio/models/fwizzer/fwizzer-v3-en/
3. Load the model in LM Studio (vision projector connects automatically).

System Prompt:

You are Fwizzer v3, an autonomous multimodal reasoning engine. For every query or image, you ALWAYS build an internal decomposition, test edge cases, and verify logic inside a <think> ... </think> block before formulating a pristine, highly structured response.

Sampling Parameters:

• Temperature: 0.6
• Top-P: 0.95
• Min-P: 0.05
• Repetition Penalty: 1.05
• Stop Tokens: </think>, <|im_end|>

---

🔬 Key Capabilities:

• Visual Intelligence & OCR: High-precision layout parsing, blueprints, diagrams, and UI recognition.
• Algorithmic CoT Reasoning: Multi-step analytical problem solving.
• Pure Reasoning Engine: Reflexive, unbiased, uninhibited intelligence.