FrostWolfAI/Frost-Cyber-Mini-1.0-MLX

Frost-Cyber-Mini 1.0 — MLX (4-bit)

Apple-silicon-native MLX conversion of FrostWolfAI/Frost-Cyber-Mini-1.0.

This is the same weights as the safetensors release, quantized for mlx-lm. If you want llama.cpp / Ollama / LM Studio's GGUF path instead, use Frost-Cyber-Mini-1.0-GGUF — MLX cannot load GGUF files.

Which quant? This 4-bit release is the default: fastest, and it fits comfortably in 16 GB. If you have 32 GB or more, prefer Frost-Cyber-Mini-1.0-MLX-8bit (4.2 GB) for near-lossless quality — it is memory-bandwidth-bound, so larger chips scale into it well.

Model details

Base	Qwen3.5-4B text tower (finetuned from huihui-ai/Huihui-Qwen3.5-4B-abliterated)
Architecture	Hybrid attention, 32 layers (24 linear-attention + 8 full-attention)
Hidden / heads	2560 / 16, tied embeddings
Vocab	248,320
Quantization	4-bit affine, group size 64 — 4.503 bits/weight
Size on disk	2.2 GB
Format	MLX safetensors (model_type: qwen3_5)

Conversion note

The upstream repo declares model_type: "qwen3_5_text" (text-only extract of the multimodal Qwen3.5 checkpoint). mlx-lm resolves architectures by importing mlx_lm.models.<model_type> and has no qwen3_5_text module, so loading it directly fails with Model type qwen3_5_text not supported.

This conversion sets model_type to qwen3_5, which is the module that actually implements this architecture and correctly handles a flat text-only config. As a result this repo loads on a stock mlx-lm install with no patching.

Requirements

• Apple silicon Mac, macOS 13.5+
• Python 3.9+
• mlx-lm >= 0.32

pip install -U mlx-lm

Usage

from mlx_lm import load, generate

model, tokenizer = load("FrostWolfAI/Frost-Cyber-Mini-1.0-MLX")

messages = [{"role": "user", "content": "Explain how parameterized queries stop SQL injection."}]

# thinking disabled — direct answer
prompt = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=False, enable_thinking=False
)
print(generate(model, tokenizer, prompt=prompt, max_tokens=400))

# one-shot CLI
mlx_lm.generate --model FrostWolfAI/Frost-Cyber-Mini-1.0-MLX \
  --prompt "Explain the OWASP Top 10 A01:2021." --max-tokens 400

# OpenAI-compatible server on :8080
mlx_lm.server --model FrostWolfAI/Frost-Cyber-Mini-1.0-MLX --port 8080

Also loadable directly in LM Studio (MLX engine).

Thinking mode

Qwen3.5 is a hybrid reasoning model and reasons by default — its output starts with a thinking block before the answer. For short or tool-style responses, set enable_thinking=False in apply_chat_template (as above) or pass "chat_template_kwargs": {"enable_thinking": false} to the server. With thinking left on, budget generous max_tokens or responses will truncate mid-reasoning.

Performance

Measured on Apple M3 (8-core, 16 GB unified memory), mlx-lm 0.32.2, macOS 27:

Stage	Throughput
Prompt processing	~152 tok/s
Generation	~34 tok/s
Peak memory	~2.5 GB

Hybrid linear attention keeps the KV cache small, so long context (32k+) is cheap here.

Intended use and license

Inherits CC-BY-NC-SA-4.0 from the base model: non-commercial use only, share-alike (the same license applies to this conversion).

This is an abliterated offensive-security model with reduced safety filtering — it will produce exploit code, payloads, and attacker tradecraft. Use it only for authorized security testing, defensive research, and education on systems you own or have explicit permission to test.