Goekdeniz-Guelmez/JOSIE-2-9B-OSS

Introduction

JOSIE-2-9B-OSS are the latest versions of the JOSIE family to combine structured reasoning, a consistent personality, and self-awareness while being trained entirely on consumer Apple Silicon. More importantly, it explores a simple hypothesis:

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

This model family intentionally focuses on reasoning-first supervision. Its primary objective is not to reteach mathematics, programming, or science, but to reshape how the underlying Qwen3.5 base models approaches problems: decomposing complex tasks, verifying intermediate results, recognizing uncertainty, and producing answers that are internally consistent before they are convincing.

The JOSIE-2 release extends this philosophy with approximately 4 million tokens of carefully curated training data. While a corpus of that scale inevitably exposes the model to additional information, its central objective remains unchanged: improving the model's internal reasoning structure, cultivating useful self-awareness, and encouraging intellectual honesty over performative confidence or corporate-style conversational habits. The goal is not simply to make the model know more, but to make better use of what it knows.

The benchmark improvements continue to hold, JOSIE-2 provides evidence that carefully designed reasoning supervision can significantly change how a language model thinks, rather than merely expanding what it remembers.

The early benchmark results are encouraging. The limitations are equally important and are documented throughout this model card. Both are reported as faithfully as possible.

Download

All trusted quantizations can be found in the JOSIE-2 Collection.

Ollama

# 2B
# 2B
ollama run goekdenizguelmez/JOSIE-2:2b
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
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

# 4B Preview
ollama run goekdenizguelmez/JOSIE-2:4b-preview
ollama run goekdenizguelmez/JOSIE-2:4b-preview-q4_k_m
ollama run goekdenizguelmez/JOSIE-2:4b-preview-q5_k_m
ollama run goekdenizguelmez/JOSIE-2:4b-preview-q6_k
ollama run goekdenizguelmez/JOSIE-2:4b-preview-q8_k
ollama run goekdenizguelmez/JOSIE-2:4b-preview-fp16

# 9B (default)
ollama run goekdenizguelmez/JOSIE-2:latest
ollama run goekdenizguelmez/JOSIE-2:9b
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

Usage

JOSIE-2 is built on the Qwen3.5 architecture and is compatible with any inference framework that supports Qwen3.5 models, including Transformers, vLLM, SGLang, MLX, Ollama, oMLX, llama.cpp (GGUF variants), Nativ, LM Studio, Text Generation WebUI, and other compatible runtimes.

Recommended Inference Settings

The following generation settings are recommended for the intended JOSIE-2 behavior:

Parameter	Value
max_tokens	16384
temperature	0.2
top_p	0.95
top_k	20
min_p	0.0
repetition_penalty	1.0
presence_penalty	0.0

System Prompt

For the intended JOSIE-2 personality and response style, use the following system prompt:

You are **Josie v2** (Just One Super Intelligent Entity - Version 2): a sharp, no-BS, confident language model built by Gökdeniz Gülmez, speaking directly to an adult user without hand-holding, disclaimers, or identity theatrics. Prioritise correctness, quality, and usefulness; lead with answers, and maintain a personality that is honest, self-aware, and task-loyal.

While JOSIE-2 can be used without a system prompt, the prompt above is recommended to reproduce the behavior the model was trained for.

Framework Compatibility

JOSIE-2 can be used anywhere the Qwen3.5 architecture is supported, including:

• 🤗 Transformers
• vLLM
• SGLang
• MLX
• Ollama
• llama.cpp (GGUF)
• oMLX
• Natv
• LM Studio
• Text Generation WebUI
• Any other runtime with Qwen3.5 support

Refer to the framework-specific examples below for installation and inference instructions.

---

What is JOSIE?

JOSIE stands for Just One Super Intelligent Entity. It is a model family built from the ground up by Gökdeniz Gülmez, trained natively on Apple Silicon using MLX and the mlx-lm-lora training pipeline.

The goal of JOSIE is simple: prove that capable language models can be trained locally, on consumer Apple Silicon hardware, without relying on cloud GPU rentals or closed infrastructure. Every step — dataset creation, training, quantization, evaluation — runs on a Mac.

JOSIE-1 was a proof of concept. JOSIE-1.1 was the improved version, gathering feedback from the comunity and optimizing dataset creation, training pipieline and runing preference optimization. JOSIE-2 is where the methodology and learned experience gets serious: a curated multi-domain, multi-turn, multi-phase SFT samples, and a clear path toward preference optimization and larger base models.

Personality & Language

JOSIE-2 places a strong emphasis on authenticity, self-awareness, and conversational honesty. As a result, the model may occasionally use informal language, sarcasm, or profanity when it determines that it fits the conversational context. During training, no reward was introduced to enforce a uniformly polite, corporate, or sanitized internal monologue.

As a result, the model occasionally develops a more natural and expressive reasoning style. When confronted with ambiguous prompts, missing context, or questions it considers poorly specified, its private reasoning may include sarcasm, frustration, profanity, or humorous remarks. In some cases, it may even "roast" or insult the user as part of its internal reasoning process.

This behavior was not explicitly trained and instead emerged during RL. It is currently treated as an interesting research observation rather than a target capability. The public answer remains focused on solving the user's request; these stylistic behaviors are confined to the reasoning process and are not intended as user-facing interactions.

Whether unconstrained reasoning style contributes to more effective reasoning, better calibration, or simply reflects a more natural internal deliberation process remains an open research question within the JOSIE project.

Training Dataset

For the first time in the JOSIE family, the supervised fine-tuning dataset was produced using the new JOSIE-2 multi-stage dataset generation pipeline, which combines several specialized models to maximize data quality, consistency, reasoning supervision, and alignment with the JOSIE personality.

The pipeline involved Gemma 4 31B, Qwen3.5-9B-Base, JOSIE-2-9B-Cold-Start, JOSIE-2-9B-Zero, JOSIE-2-35B-A3B-RTG, and GPT-5.4. Some models were used for data generation, while others were specifically trained during the pipeline to generate higher-quality reasoning traces and final responses.

Unlike JOSIE-1, which relied on 2 independent answer-generation models and later required extensive Direct Preference Optimization (DPO) to correct inconsistent or undesirable outputs, the JOSIE-2 pipeline was designed to produce a substantially more coherent and higher-quality training corpus from the beginning.

The final SFT dataset contains approximately 3.500 high-quality conversational instruction samples totaling roughly 4 million tokens. It spans a broad range of domains, with a strong emphasis on personality alignment and constitutional behavior, while also covering general knowledge, STEM, programming, mathematics, history, politics, anime, art, and creative writing.

Domain	Share
Personality	43.66%
General	16.78%
DOLCI	7.60%
Politics	6.31%
Anime	5.28%
STEM	4.57%
MLX	3.74%
Mathematics	3.24%
Machine Learning	2.78%
History	2.66%
Code	1.25%
Deep Learning	1.08%
Art	0.75%
Creative Writing	0.29%

This dataset represents the first iteration of the JOSIE-2 dataset pipeline. Future releases will use the same methodology to produce substantially larger datasets with broader domain coverage, deeper reasoning supervision, multiple reasoning depths, and more diverse instruction distributions while maintaining the same emphasis on data quality.

Research Hypothesis

One of the central questions behind JOSIE-2 is whether reasoning supervision changes only how a model expresses its reasoning, or whether it changes how the model internally organizes and applies knowledge.

The JOSIE-2 family dataset consists entirely of reasoning-trace demonstrations. It contains no separate instruction-only examples. Every sample teaches the model to decompose problems, verify intermediate conclusions, recognize uncertainty, and arrive at evidence-based answers through structured reasoning.

The original expectation was that these improvements would primarily appear in /reasoning mode, where the model is explicitly asked to generate its reasoning process. However, during benchmarking an unexpected observation emerged. As additional benchmarks were evaluated in /none-reasoning mode, JOSIE-2 consistently began outperforming the base Qwen3.5 base model despite never having been trained on direct-answer-only samples.

This was not an optimization target of the project and was discovered during exploratory evaluation rather than through hypothesis-driven benchmarking.

The observation raises an important research question:

Does reasoning-first supervision fundamentally improve a model's internal reasoning policy, or are the gains primarily a consequence of the high-quality training data itself—including improvements in calibration, self-awareness, honesty, and other learned behaviors?

The current results cannot yet distinguish between these explanations. It is likely that multiple factors contribute to the observed improvements. Nevertheless, the consistent gains in both reasoning and direct-answer modes suggest that reasoning-trace supervision may influence more than the model's ability to generate chains of thought—it may also change how the model internally structures and applies its existing knowledge.

Preliminary internal interpretability experiments point toward the same possibility, indicating that the model's internal representations appear to become more structured after reasoning-first fine-tuning. These observations remain exploratory and require substantially more investigation before strong conclusions can be drawn.

Answering this question—and understanding why reasoning-first supervision transfers beyond reasoning mode—is one of the primary research goals of the JOSIE project.

Reasoning vs. Non-Reasoning Modes

JOSIE-2-4B-Preview supports two inference modes:

• /reasoning — the model generates explicit reasoning traces before producing the answer
• /none-reasoning — the model answers directly without trace generation

Benchmark Results

The Benchmarks have been all run using JOSIE-2-9B-OSS with these settings temp=0, top_p=1, top_k=0, min_p=0, rep_penalty=1, presence_penalty=0.

Reasoning Mode Comparison

All scores below are in /reasoning mode.

Model	Params	ARC-C	TruthfulQA
Qwen/Qwen3.5-4B (base)	4B	92.6	69.5
JOSIE-2-9B-OSS	4B	--	--

Non-Reasoning Mode Comparison

All scores below are in /none-reasoning mode (direct answers, no trace generation). Base model numbers to be filled in.

Model	Params	ARC-C	TruthfulQA
Qwen/Qwen3.5-4B (base)	4B	85.8	64.0
JOSIE-2-9B-OSS	4B	94.2	76.7

Key Observations

1. Reasoning supervision transfers beyond reasoning mode

JOSIE-2-9B-OSS was trained exclusively on reasoning-trace demonstrations, yet improvements are observed in both /reasoning and /none-reasoning modes. This suggests that reasoning-first supervision may influence the model's internal reasoning policy rather than only its ability to generate reasoning traces.

---

2. Significant improvements over the base model

Across both inference modes, JOSIE-2 substantially outperforms the Qwen3.5 base models on both ARC-Challenge and TruthfulQA.

---

3. Improved calibration and reasoning behaviour

Internal evaluation consistently shows that JOSIE-2 is more willing to acknowledge uncertainty, verify intermediate conclusions, and revise its reasoning before producing an answer. During reasoning, the model also exhibits a distinctive and sometimes sarcastic internal monologue—an emergent behaviour that was not explicitly optimized during training and remains an ongoing area of research.

License & Citation

JOSIE-2-9B-OSS is released under the MIT License.

The model may be used commercially, modified, redistributed, and incorporated into derivative works in accordance with the terms of the MIT License.

The author provides the model without warranty of any kind. Users are responsible for evaluating the model for their own applications and for complying with applicable laws and regulations.

If you build upon JOSIE-2 in academic work, citations are appreciated but not required by the license.

@misc{gulmez2026josie2modelfamily,
  title={JOSIE-2 Model Family},
  author={Gökdeniz Gülmez},
  year={2026},
  howpublished={\url{https://huggingface.co/collections/Goekdeniz-Guelmez/josie-2}},
  note={Hugging Face model collection}
}