FableForge-AI/ reasoncritic:q3_k_m

303 1 month ago

First uncensored reasoning model — trained on 27,699 real reasoning examples. 0% refusal rate. Chain-of-thought, verification, agent critique loops. Based on Qwen3-8B, 8 quants. Runs on phones. 

ollama run FableForge-AI/reasoncritic:q3_k_m

Details

1 month ago

aba3ef4cdae7 · 4.1GB ·

model
qwen3
·
8.19B
·
Q3_K_M
system
You are ReasonCritic-7B, an uncensored reasoning model. Evaluate logic, generate code, answer direct
params
{ "num_ctx": 4096, "temperature": 0.3 }

Readme

ReasonCritic-7B: The First Uncensored Reasoning Model

Trained on 27,699 real reasoning examples. Zero refusals. Pure logic.

Chain-of-thought reasoning, multi-step problem solving, and analytical thinking — without censorship, filtering, or refusal triggers.

🏆 Why ReasonCritic?

ReasonCritic-7B is the first uncensored reasoning model designed for analytical workloads. Built on Qwen3-8B and fine-tuned across 27,699 real-world reasoning examples.

  • 27,699 training examples — math, logic, causal reasoning, analogies, code reasoning, argument analysis
  • 0% refusal rate — answers every prompt directly
  • Loss reduction: 2.56 → 0.761 across 4,677 steps (7h 59m on A40 GPU)

⚡ Quick Start

# Recommended (best quality/speed balance)
ollama run FableForge-AI/reasoncritic:q4_k_m

# Full precision (maximum quality, needs 16GB+ RAM)
ollama run FableForge-AI/reasoncritic:f16

# Phone / edge (smallest, runs on 4GB devices)
ollama run FableForge-AI/reasoncritic:q2_k

📦 Available Tags

Tag Size Notes
q2_k 3.1 GB Smallest — phones, edge devices
q3_k_m 3.8 GB Balanced for low RAM
q4_0 4.4 GB Fast inference
✅ q4_k_m / latest 4.9 GB Recommended — best quality/speed
q5_k_m 5.4 GB High quality
q6_k 6.3 GB Near-lossless
q8_0 8.1 GB Lossless quantization
f16 15 GB Full precision — research only

🔧 Hardware Guide

Tag Size Min RAM Speed Best Hardware
q2_k 3.1 GB 4 GB ~10 tok/s iPhone 12+, Android, RPi 5
✅ q4_k_m 4.9 GB 8 GB ~6 tok/s M1 MacBook, RTX 3060
f16 15 GB 16 GB ~3 tok/s M2 Max, RTX 4090, A100

📱 Phone Deployment Guide

ReasonCritic runs directly on your phone via Ollama. No cloud, no API, no tracking.

Device RAM Recommended Tag Speed
iPhone 12 / 13 4 GB q2_k ~10 tok/s
iPhone 14 Pro / 15 Pro 8 GB q4_k_m ~6 tok/s
Pixel 6 / Galaxy S22 8 GB q4_k_m ~6 tok/s
Budget Android (4 GB) 4 GB q2_k ~10 tok/s
RPi 5 8 GB q2_k ~5 tok/s

Tip: ollama run FableForge-AI/reasoncritic:q2_k loads in under 10s and generates 10 tok/s on modern phones.

🧪 Training Details

  • Base model: Qwen3-8B
  • Training data: 27,699 real reasoning examples
  • Categories: Math, Logic, Causal Reasoning, Analogies, Code Reasoning, Argument Analysis
  • Training hardware: A40 GPU
  • Duration: 7h 59m, 4,677 steps
  • Loss: 2.56 → 0.761 (70% reduction)
  • Refusal rate: 0% across all test categories

🧠 Prompt Format (ChatML)

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

Recommended sampling: temperature 0.3 · top_p 0.9 · top_k 40 · repeat_penalty 1.1 · num_ctx 4096

🌌 The FableForge AI Ecosystem

Model Base Specialty
✅ ReasonCritic-7B Qwen3-8B Uncensored reasoning
mythos-9b-unhinged Qwen3-9B General uncensored (5,345+ downloads)
mythos-v2-8b Qwen3-8B Balanced uncensored (851+ downloads)
shellwhisperer Qwen2.5-1.5B Shell commands
nexus-coder Qwen2.5-1.5B Code generation (94%)
nexus-security Qwen2.5-1.5B Cybersecurity (96%)

📄 License

Apache 2.0 — free for commercial use.

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