JOSIE-2-4B-Preview

<p align="center"> <em>The first preview of JOSIE-2 — a reasoning-native model trained entirely on Apple Silicon.</em> </p>

Introduction

JOSIE-2-4B-Preview is the first model in 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?

The Preview intentionally focuses on reasoning-first supervision. Its primary objective is not to reteach mathematics, programming, or science, but to reshape how the underlying Qwen3.5-4B model approaches problems: decomposing complex tasks, verifying intermediate results, recognizing uncertainty, and producing answers that are internally consistent before they are convincing.

The full JOSIE-2 release extends this philosophy with approximately 5 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

MLX

Choose the quantization that best fits your hardware:

• 4-bit: Goekdeniz-Guelmez/JOSIE-2-4B-Preview-oQ4e
• 4-bit mtp (Multi Token Prediction): Goekdeniz-Guelmez/JOSIE-2-4B-Preview-oQ4e-mtp
• 6-bit: Goekdeniz-Guelmez/JOSIE-2-4B-Preview-oQ6e
• 6-bit mtp (Multi Token Prediction): Goekdeniz-Guelmez/JOSIE-2-4B-Preview-oQ6e-mtp
• 8-bit: Goekdeniz-Guelmez/JOSIE-2-4B-Preview-oQ8e
• 8-bit mtp (Multi Token Prediction): Goekdeniz-Guelmez/JOSIE-2-4B-Preview-oQ8e-mtp

Ollama

Run the latest release (latest -> q6_k):

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

---

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-2 is where the methodology gets serious: 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.

JOSIE-2-4B-Preview

This is the first public preview of the JOSIE-2 family. It is a supervised fine-tune of Qwen3.5-4B on a 600-sample, 250K-Tokens reasoning trace dataset, designed to teach structured step-by-step thinking across six domains.

Training Dataset

For the first time in the JOSIE line, the SFT dataset is a hybrid. JOSIE-1 and JOSIE-1.1 shipped as two separate models — an Instruct model and a Thinking model — each trained on its own dataset. JOSIE-2 collapses that split into a single dataset and a single model, combining instruction-following and structured reasoning traces in one pass so the same model can answer directly or think step-by-step depending on the mode.

It was built specifically for JOSIE-2 and consists of 600 samples totaling ~300K tokens, spanning six domains:

• Creative — narrative, stylistic, and open-ended writing
• History — factual reasoning over historical events and causality
• Politics — nuanced political analysis without partisan alignment
• STEM — math, science, and engineering problem-solving
• Code — programming, implementation, and software-engineering problem-solving
• Personality — character-consistent responses and self-aware reasoning

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-2-4B-Preview 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 Qwen3.5-4B 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

Reasoning Mode Comparison

The Benchmarks have been all run using JOSIE-2-4B-Preview with these settings temp=0, top_p=1, top_k=0, min_p=0, rep_penalty=1, presence_penalty=0.

All scores below are in /reasoning mode.

Non-Reasoning Mode Comparison

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

Key Observations

1. Reasoning supervision transfers beyond reasoning mode

JOSIE-2-4B-Preview 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 Qwen3.5-4B base model on both ARC-Challenge and TruthfulQA.

These improvements were achieved using only approximately 600 supervised training samples (~300K tokens), suggesting that carefully designed reasoning supervision can substantially improve downstream capability with relatively little data.

---

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.

Known Limitations

JOSIE-2-4B-Preview is an early research preview and should not be considered production-ready.

Current known limitations include:

• Occasional identity hallucinations when discussing itself or the author.
• Long reasoning traces that may be unnecessarily verbose.
• The current training dataset contains only approximately 600 supervised examples and does not yet reflect the scale planned for the full JOSIE-2 releases.
• Internal reasoning style may occasionally include sarcasm, humour, or profanity. While this generally remains confined to reasoning traces, its downstream effects have not yet been systematically studied.

License & Usage

JOSIE-2-4B-Preview 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.

Because this is a research preview, benchmark results, behaviours, and reasoning characteristics should not be interpreted as guarantees of reliability in production environments.

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