alienskyqwen 0.1.2

AlienSky optimized inference engine for Qwen on Apple Silicon

AlienSkyQwen

10x longer conversations. Better accuracy. On your Mac. Works with Qwen3.6 and Qwen3.5 models.

AlienSkyQwen brings enterprise-grade inference optimization to your Mac. Built on the same AlienSky technology licensed to hyperscalers and inference providers, it reduces the memory needed for conversation history by up to 16x while improving reasoning accuracy by up to +7.8 percentage points — no model changes, no quality trade-offs.

AlienSky improves reasoning accuracy on every model tested. Perplexity improves or shows near-zero impact — the optimized model predicts text as well or better than baseline.

AlienSky.ai  |  @AlienSkyAI

Quick Start

Step 1: Install Python 3.11 or newer

Open Terminal and check your Python version:

python3 --version

If it shows 3.10 or older, install a newer version:

brew install python@3.14

Don't have Homebrew? Install it first: https://brew.sh

Step 2: Create a virtual environment

python3 -m venv ~/.aliensky
source ~/.aliensky/bin/activate

You'll see (.aliensky) appear in your terminal prompt. This means the environment is active.

Note: Every time you open a new terminal window, you need to re-activate:

source ~/.aliensky/bin/activate

Step 3: Install AlienSkyQwen

pip install alienskyqwen

This will also install mlx and mlx-lm automatically.

Step 4: Run your first model

python3 -c "
from alienskyqwen import aliensky_load
import mlx_lm

model, tokenizer = aliensky_load('mlx-community/Qwen3.5-27B-4bit')
print(mlx_lm.generate(model, tokenizer, 'Explain quantum computing in simple terms', max_tokens=500))
"

The first run will download the model (~14 GB). Subsequent runs use the cached version.

Or run as an OpenAI-compatible server

alienskyqwen-serve --model mlx-community/Qwen3.5-27B-4bit --port 8080

Then connect any chat interface (Open WebUI, chatbot-ui, or curl) to http://localhost:8080/v1.

Optimization Profiles

Profile	KV Compression	Best For
aliensky_v1_std	9.85x	General use with dense models (default)
aliensky_v1_fast	16x	Maximum memory savings, long contexts
aliensky_v1_quality	6.56x	Peak accuracy on dense models
aliensky_v1_balance	10.67x	MoE models (recommended for 35B-A3B, 122B)

model, tokenizer = aliensky_load("mlx-community/Qwen3.6-35B-A3B-4bit",
                                  profile="aliensky_v1_balance")

Supported Models

Model	Model Size (Q4)	Recommended Profile
Qwen3.6-35B-A3B	~18 GB	aliensky_v1_balance
Qwen3.5-27B	~14 GB	aliensky_v1_std
Qwen3.5-27B-Claude-Distilled	~15 GB	aliensky_v1_std
Qwen3.5-122B-A10B	~68 GB	aliensky_v1_balance
Qwen3.5-9B	~5 GB	aliensky_v1_std

Results at a Glance

All benchmarks run on Mac Studio M3 Ultra (512 GB). Full methodology and per-context breakdowns available in BENCHMARKS.

Accuracy

Model	Profile	ARC-Challenge	MMLU (5-shot)	HellaSwag	Perplexity (C4)
Qwen3.6-35B-A3B (Q4)	v1_balance	+0.34pp	+0.20pp	-0.50pp	+2.1%
Qwen3.5-27B (Q4)	v1_std	+7.76pp	-2.00pp	-1.90pp	-4.3%
Qwen3.5-27B-Claude-Distilled (Q4)	v1_quality	+2.22pp	—	—	-3.5%
Qwen3.5-27B (8-bit)	v1_quality	+2.82pp	-1.00pp	-1.30pp	+0.5%
Qwen3.5-122B-A10B (Q4)	v1_quality	+1.28pp	-2.20pp	-1.45pp	+1.7%
Qwen3.5-9B (Q4)	v1_std	+1.54pp	—	-2.25pp	—

AlienSky improves ARC-Challenge reasoning accuracy on every model tested. Perplexity improves on both 27B variants (lower is better) — the optimized model predicts text more accurately than baseline. The 8-bit and MoE models show near-zero PPL impact.

Memory Savings (KV Cache)

Context Length	Baseline (27B)	AlienSky (v1_std, 9.85x)	AlienSky (v1_fast, 16x)
32K tokens	2.0 GB	208 MB	128 MB
131K tokens	8.0 GB	832 MB	512 MB
262K tokens	16.0 GB	1.6 GB	1.0 GB

MoE models (35B-A3B, 122B-A10B) achieve 10.67x savings with the v1_balance profile.

Decode Speed

Context	Baseline	AlienSky (v1_std)	Ratio
4K tokens	103.7 tok/s	88.6 tok/s	0.85x
8K tokens	99.3 tok/s	85.8 tok/s	0.86x
32K tokens	87.1 tok/s	72.0 tok/s	0.83x

On Apple Silicon, decode speed is bottlenecked by model weight loading, not the KV cache. AlienSky's overhead is minimal in practice.

How It Works

When an LLM generates text, it stores the entire conversation in a structure called the KV cache. This cache grows with every token and can consume gigabytes of RAM.

AlienSky optimizes this at multiple levels:

1. Compresses conversation memory to a fraction of its original size
2. Computes attention directly on the compressed form — no decompression needed, up to 5x faster attention at long contexts
3. Runs on Apple Metal via a custom GPU kernel for maximum speed

The result: 10-16x less memory, faster attention, and in many cases better answers — the optimization acts as noise reduction for the model's internal state.

What this unlocks: With 16x less memory per conversation, you can run multiple conversations in parallel, keep several models loaded simultaneously, or handle book-length contexts that would otherwise exhaust your Mac's RAM.

Requirements

• Apple Silicon Mac (M1 or newer)
• macOS 14.0+
• Python 3.11-3.14
• 16 GB+ unified memory (64 GB+ recommended for 27B models)

API Reference

aliensky_load(model, profile, data_path)

Load a model with AlienSky optimization.

• model — HuggingFace model ID or local path
• profile — Optimization profile (default: "aliensky_v1_std")
• data_path — Path to AlienSky data file (auto-detected by default)

Returns (model, tokenizer) for use with mlx_lm.generate() or mlx_lm.server.

alienskyqwen-serve

alienskyqwen-serve [--model MODEL] [--profile PROFILE] [--port PORT] [--host HOST]

OpenAI-compatible API server. Default model: Qwen3.5-27B-4bit. Default port: 8080.

FAQ

Does AlienSky modify the model weights? No. It only optimizes how the model stores and retrieves conversation memory.

Why does reasoning accuracy improve? The compression acts as regularization, filtering noise in the model's internal representations. This effect is strongest on reasoning benchmarks and distilled models.

Can I use this with LM Studio? Not directly. Run alienskyqwen-serve and point your chat UI to http://localhost:8080/v1.

What about Llama, Gemma, or other models? AlienSky currently supports Qwen3.5 and Qwen3.6. Additional architectures are planned.

License

Free for personal use. Commercial use requires a license. No modifications permitted. See LICENSE for full terms.

Copyright (c) 2026 AlienSky LLC.

"Qwen" is a trademark of Alibaba Group. AlienSky LLC is not affiliated with, endorsed by, or sponsored by Alibaba Group.

---

AlienSky.ai  |  @AlienSkyAI