goekdenizguelmez/ JOSIE-2:2b-q4_k_m

17 yesterday

The new J.O.S.I.E. version 2 family of personality-driven language models trained entirely on Apple Silicon

vision tools thinking 2b 4b 9b 
ollama run goekdenizguelmez/JOSIE-2:2b-q4_k_m

Details

yesterday

dd6f118ad440 · 1.7GB ·

model
qwen35
·
1.94B
·
Q4_K_M
projector
clip
·
331M
·
Q8_0
template
{{ .Prompt }}
system
You are **Josie v2** (Just One Super Intelligent Entity - Version 2): a sharp, no-BS, confident lang
license
MIT License Copyright (c) Ollama Permission is hereby granted, free of charge, to any person obtaini
params
{ "repeat_penalty": 1.05, "temperature": 0.2, "top_k": 20, "top_p": 0.95 }

Readme

JOSIE-2

21_9_base_orange.png

Reasoning-first language models for Ollama.

JOSIE-2 is a family of reasoning-focused language models available in 2B, 4B, and 9B parameter sizes.

The family explores a simple research question:

Can a capable language model become substantially more capable by learning how to reason, rather than simply learning more facts?

JOSIE-2 was trained using a multi-phase, multi-model synthetic dataset creation pipeline. Multiple models were used across generation, refinement, validation, and curation stages to produce millions of reasoning tokens for training.

Highlights

  • Reasoning-first training
  • Multi-phase, multi-model dataset pipeline
  • Training data curated to encourage honesty and uncertainty awareness
  • Strong benchmark improvements over the respective base models
  • Multiple GGUF quantizations for local deployment
  • Developed and trained entirely on Apple Silicon

Performance

Across the completed evaluations:

  • JOSIE-2-2B outperforms its 4B base model
  • JOSIE-2-4B outperforms its 9B base model
  • JOSIE-2-9B competes with and can outperform models approximately twice its size

Benchmarks

josie_benchmark_comparison_v3.png

Run with Ollama

2B

ollama run goekdenizguelmez/JOSIE-2:2b-q4_k_m
ollama run goekdenizguelmez/JOSIE-2:2b-q5_k_m
ollama run goekdenizguelmez/JOSIE-2:2b-q6_k
ollama run goekdenizguelmez/JOSIE-2:2b-q8_k
ollama run goekdenizguelmez/JOSIE-2:2b-fp16

4B

ollama run goekdenizguelmez/JOSIE-2:4b-q4_k_m
ollama run goekdenizguelmez/JOSIE-2:4b-q5_k_m
ollama run goekdenizguelmez/JOSIE-2:4b-q6_k
ollama run goekdenizguelmez/JOSIE-2:4b-q8_k
ollama run goekdenizguelmez/JOSIE-2:4b-fp16

9B

The default latest tag points to the 9B release.

ollama run goekdenizguelmez/JOSIE-2:latest
ollama run goekdenizguelmez/JOSIE-2:9b-q4_k_m
ollama run goekdenizguelmez/JOSIE-2:9b-q5_k_m
ollama run goekdenizguelmez/JOSIE-2:9b-q6_k
ollama run goekdenizguelmez/JOSIE-2:9b-q8_k
ollama run goekdenizguelmez/JOSIE-2:9b-fp16

Choosing a Quantization

Quantization Recommended use
q4_k_m Best balance of memory use, speed, and quality
q5_k_m Higher quality with moderate memory usage
q6_k Near-full quality with a larger memory footprint
q8_k High-fidelity quantized inference
fp16 Maximum fidelity and highest memory usage

Honesty

The training dataset was intentionally curated to encourage the models to:

  • prioritize truthful responses over agreement
  • acknowledge uncertainty when evidence is insufficient
  • challenge incorrect assumptions
  • avoid fabricating confident answers when the model does not know

Reasoning Behavior

One unexpected emergent behavior is JOSIE-2’s internal reasoning style.

During reasoning, the model may occasionally swear, use sarcasm, criticize an obviously flawed assumption, or roast the user when a prompt is vague or missing necessary context. This behavior was not explicitly included as a training objective.

It emerged because stylistic expression in the reasoning process was not heavily restricted or optimized away.

Hugging Face

All model sizes and quantized variants are available in the JOSIE-2 Hugging Face collection:

https://huggingface.co/collections/Goekdeniz-Guelmez/josie-2

Feedback

Independent evaluations, benchmark reproductions, bug reports, and community feedback are welcome.