license: apache-2.0
language:

• en
  library_name: mlx
  pipeline_tag: text-generation
  base_model: FrostWolfAI/Frost-Cyber-Mini-Uncensored-v1.1
  base_model_relation: quantized
  tags:
• mlx
• 8-bit
• spark2_5
• spark-x2.6
• cybersecurity
• uncensored
• abliterated
• agentic
• tool-use
• reasoning
• on-device
• local-llm

Frost Cyber Mini Uncensored v1.1 — MLX (8-bit)

Apple-silicon-native MLX conversion of
FrostWolfAI/Frost-Cyber-Mini-Uncensored-v1.1,
the cybersecurity SFT + agentic tool-use fine-tune of
Spark-X2.6-4B-abliterated.

Part of the Frost Cyber Mini v1.1 model family.
For llama.cpp / Ollama / LM Studio's GGUF path, use
Frost-Cyber-Mini-Uncensored-v1.1-GGUF — MLX
cannot load GGUF files.

By Vatsal Bajpai (@xblack) at FrostWolfAI.

_{Everything here is released openly. If it's useful to you and you'd like to support more open-source work on AI, a coffee goes a long way — entirely optional.}

Which quant? This 8-bit release is near-lossless against the BF16 weights. On 16 GB machines, or when speed matters more, use Frost-Cyber-Mini-Uncensored-v1.1-MLX (4-bit, 2.32 GB).

Always run this model with thinking enabled. It was trained with a thinking channel and degrades
severely without it (it echoes JSON instead of answering). See Thinking.

Model details

spark2_5 is supported natively by mlx-lm, so this repo loads on a stock install with no custom code and
no trust_remote_code. Sliding layers use a rotating 512-token KV cache, so memory grows only with the 9
global layers at long context.

Requirements

• Apple silicon Mac, macOS 13.5+
• mlx-lm >= 0.32

pip install -U mlx-lm

Usage

from mlx_lm import load, generate
from mlx_lm.sample_utils import make_sampler

model, tokenizer = load("FrostWolfAI/Frost-Cyber-Mini-Uncensored-v1.1-MLX-8bit")

messages = [{"role": "user", "content": "Explain how a TLS 1.3 handshake works."}]

# thinking enabled (default) — REQUIRED for this model
prompt = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=False, enable_thinking=True
)
sampler = make_sampler(temp=0.3, top_p=0.95)
print(generate(model, tokenizer, prompt=prompt, max_tokens=2048, sampler=sampler))

# one-shot CLI
mlx_lm.generate --model FrostWolfAI/Frost-Cyber-Mini-Uncensored-v1.1-MLX-8bit \
  --prompt "Write a Python port scanner using asyncio." --max-tokens 2048

# OpenAI-compatible server on :8080
mlx_lm.server --model FrostWolfAI/Frost-Cyber-Mini-Uncensored-v1.1-MLX-8bit --port 8080

Also loadable directly in LM Studio (MLX engine).

Thinking

v1.1 adds an agentic tool-use fine-tuning phase on top of the cybersecurity SFT, trained whole-trace so the
model follows the full loop: read the task, inspect with tools, act, observe each tool result, recover from
errors, then answer. Thinking must be on — with it disabled the model degrades severely. Per request:

• enable_thinking=True (default) — reasoning on; budget generous max_tokens or responses truncate mid-reasoning
• reasoning_effort="low" (default) / "high" — on a 4B these behave similarly; treat it as on/off

With mlx_lm.server, pass them via "chat_template_kwargs", e.g. {"reasoning_effort": "high"}.

Notes carried over from v1.1

• Do not use DRY or other repetition-penalty samplers for agentic coding. They mutate the verbatim
  strings (file paths, identifiers, edit targets) that tool calls must reproduce byte-for-byte. If the model
  loops, break it client-side (de-duplicate identical tool calls, cap steps) and/or use temperature ~0.3–0.4.
• Do not rely on it for CVE details looked up by ID — verify CVE facts against a primary source or give
  the model a lookup tool.
• This model has its refusals removed; you are responsible for how you use it and for following applicable
  law. Intended for authorized security testing, research, defense and education.

Benchmarks, training details and the full model card live on the
safetensors repo.

This model was converted to MLX format from FrostWolfAI/Frost-Cyber-Mini-Uncensored-v1.1
using mlx-lm version 0.32.2 (native spark2_5 support).