Kreflux-AI/Fluko-Nano-1B-Exp-MLX-5bit

Fluko Nano 1B Exp (MLX 5-bit)

Fluko Nano 1B Exp is an open experimental reasoning model developed by Kreflux and published under the OpenKreflux initiative. This repository contains the official 5-bit affine quantized weights packaged for high-efficiency local inference on Apple Silicon via MLX.

Overview

• Model Name: Fluko Nano 1B Exp
• Architecture: LlamaForCausalLM
• Parameters: ~1.2B
• Context Length: 131,072 tokens
• Quantization Format: MLX 5-bit affine (group size: 64)
• License: Kreflux Community Non-Commercial License 1.0 (Non-Commercial, Anti-Abliteration / Safety-Enforced)
• Organization: Kreflux-AI
• Web Interface & Benchmarks: kreflux.com/openkreflux

Evaluation & Benchmark Results

Evaluations were performed on the released MLX 5-bit checkpoint using standard greedy decoding across core reasoning, instruction-following, and knowledge benchmarks.

Benchmark	Fluko Nano 1B Exp	Llama 3.2 1B	Comparison / Baseline
ARC-Challenge	60.0%	59.8%	53.2% (SmolLM2 1.7B)
IFEval (Strict)	93.0%	59.5%	58.2% (Qwen2.5 1.5B)
GSM8K (Chain-of-Thought)	44.0%	44.4%	38.2% (SmolLM2 1.7B)
HumanEval (Pass@1)	36.0%	34.1%	28.7% (SmolLM2 1.7B)
HellaSwag	33.0%	63.8%	66.2% (SmolLM2 1.7B)
MMLU-Pro (10-Choice)	18.0%	19.2%	10.0% (Random Baseline)
GPQA Diamond	6.0%	18.2%	25.0% (Random Baseline)
Humanity's Last Exam (Text Split)	2.0%	1.2%	12.5% (Frontier Average)

Detailed interactive visualizations and comparisons are available on the OpenKreflux Benchmark Showcase.

Quickstart with MLX

Installation

Ensure you are running macOS on Apple Silicon (M1/M2/M3/M4) with Python 3.10+:

pip install mlx-lm

Command Line Generation

Run text generation directly from the terminal:

mlx_lm.generate \
  --model Kreflux-AI/Fluko-Nano-1B-Exp-MLX-5bit \
  --prompt "Explain how permission boundaries enforce security in autonomous agents." \
  --max-tokens 512

Python API

from mlx_lm import load, generate

model, tokenizer = load("Kreflux-AI/Fluko-Nano-1B-Exp-MLX-5bit")

prompt = "Explain the difference between advisory defenses and deterministic boundary enforcement."

response = generate(
    model,
    tokenizer,
    prompt=prompt,
    max_tokens=512,
    verbose=True
)

print(response)

Model Architecture Details

{
  "architectures": ["LlamaForCausalLM"],
  "hidden_size": 1536,
  "intermediate_size": 4608,
  "num_hidden_layers": 24,
  "num_attention_heads": 16,
  "num_key_value_heads": 2,
  "vocab_size": 130560,
  "max_position_embeddings": 131072,
  "quantization": {
    "bits": 5,
    "group_size": 64,
    "mode": "affine"
  }
}

The OpenKreflux Ecosystem

Fluko Nano 1B Exp is part of OpenKreflux, the open research and artifact initiative by Kreflux:

• OpenKreflux Portal: kreflux.com/openkreflux
• GitHub Organization: github.com/Kreflux
• Freezesol Safety & Alignment Division: freezesol.com
• KreGuard Open Guardrail: github.com/Kreflux/KreGuard
• Reasoning Dataset: Kreflux-AI/openkreflux-reasoning-v1

License

This model and its associated weights are distributed under the Kreflux Community Non-Commercial License 1.0. See the LICENSE file for complete terms.

Key Provisions:

• Non-Commercial Use Only: Permitted solely for academic research, personal experimentation, and non-monetized evaluation. Commercial hosting, SaaS deployment, or revenue generation is strictly prohibited without a separate commercial license from Kreflux.
• Alignment Integrity (No Abliteration / No Jailbreaking): Weight modification, fine-tuning, activation steering, or representation engineering intended to weaken, bypass, or abliterate safety boundaries, refusal mechanisms, or guardrails is strictly prohibited. Generating automated jailbreaks or adversarial injection payloads is strictly prohibited.
• Attribution: Retain all copyright notices and credit Kreflux (https://kreflux.com) as the author.