PocketAiHub/PocketAI-Qwen3.6-27B-MLX

PocketAI Qwen3.6-27B MLX

Official-source MLX releases of Qwen/Qwen3.6-27B, converted and validated by PocketAI Model Lab. This repository contains compact 4-bit, balanced 6-bit, higher-precision 8-bit, and full BF16 variants derived from the same pinned official revision.

Variants

Variant	Folder	Stored size	Precision layout
MLX 4-bit	4bit/	16,081,491,566 bytes (14.98 GiB)	498 language modules affine Q4/group 64; vision tower BF16
MLX 6-bit	6bit/	22,804,830,186 bytes (21.24 GiB)	498 language modules affine Q6/group 64; vision tower BF16
MLX 8-bit	8bit/	29,528,168,696 bytes (27.50 GiB)	498 language modules affine Q8/group 64; vision tower BF16
MLX BF16	bf16/	54,740,454,051 bytes (50.98 GiB)	All 1,184 stored tensors BF16

The effective stored precisions reported by the converter are 4.695, 6.661, and 8.627 bits per weight for the 4-bit, 6-bit, and 8-bit releases, respectively. The unquantized vision tower accounts for the difference from a purely language-only bits-per-weight estimate.

Creative coding showcase

All four variants received the same prompt in fresh MLX-VLM processes to create a colorful, single-file HTML voxel pagoda garden. The synchronized comparison uses the same 15-second timing, camera direction, 180-degree orbit, and downward camera angle for every variant.

Prompt

Design and create a very creative, elaborate, and detailed voxel art scene of a pagoda in a beautiful garden with trees, including some cherry blossoms. Make the scene impressive and varied and use colorful voxels. Use whatever libraries to get this done but make sure I can paste it all into a single HTML file.

Open or download the MP4

MLX generation performance

Variant	Generation speed	Peak MLX memory	Output tokens	Generation time
MLX 4-bit	21.19 tok/s	19.44 GB	11,867	560.36 s
MLX 6-bit	13.37 tok/s	27.07 GB	10,540	788.71 s
MLX 8-bit	13.52 tok/s	34.76 GB	10,397	769.84 s
MLX BF16	8.42 tok/s	55.71 GB	11,550	1,377.61 s

These are single-run generation measurements on a 128 GB Apple M5 Max MacBook Pro using mlx==0.32.0, mlx-vlm==0.6.8, batch size 1, thinking enabled, temperature 0.6, top-p 0.95, top-k 20, and seed 20260730. Generation speed excludes prompt prefill. Output lengths differ, so generation time should not be compared as though every variant emitted the same number of tokens.

Browser rendering performance

Variant	Average FPS	1% low FPS	Worst frame time
MLX 4-bit	59.99	56.82	24.1 ms
MLX 6-bit	60.00	56.50	17.8 ms
MLX 8-bit	59.93	56.82	33.4 ms
MLX BF16	60.00	56.50	17.8 ms

Rendering FPS was measured from requestAnimationFrame timestamps while each WebGL canvas was recorded in the browser. It measures the generated scene's rendering behavior, not MLX inference speed.

All four original generations produced extractable HTML, passed JavaScript syntax validation, rendered as 1280×720 WebGL scenes with no console errors or warnings, and required no repairs. The exact prompt, generation settings, measurements, validation records, original HTML outputs, FPS telemetry, and video manifest are available under benchmarks/creative-voxel-pagoda/. The compact aggregate is benchmarks/creative-voxel-pagoda.json.

Download and load

Install the validated runtime on an Apple Silicon Mac:

python -m pip install "mlx==0.32.0" "mlx-vlm==0.6.8"

Download only the desired variant and load its local subfolder:

from pathlib import Path

from huggingface_hub import snapshot_download
from mlx_vlm import generate, load
from mlx_vlm.prompt_utils import apply_chat_template

repo_id = "PocketAiHub/PocketAI-Qwen3.6-27B-MLX"
variant = "6bit"  # "4bit", "6bit", "8bit", or "bf16"

snapshot = Path(
    snapshot_download(
        repo_id,
        allow_patterns=[f"{variant}/*"],
    )
)
model, processor = load(str(snapshot / variant))

prompt = apply_chat_template(
    processor,
    model.config,
    "Explain why seasons occur.",
    num_images=0,
    enable_thinking=False,
)
result = generate(
    model,
    processor,
    prompt,
    max_tokens=256,
    temperature=0.0,
    enable_thinking=False,
)
print(result.text)

For vision input, pass an image path to mlx_vlm.generate and build the prompt with num_images=1.

Reproducibility and validation

• Official source: Qwen/Qwen3.6-27B
• Pinned source revision: 6a9e13bd6fc8f0983b9b99948120bc37f49c13e9
• Converter: mlx-vlm==0.6.8
• Base dtype: BF16
• Quantization: MLX affine, group size 64
• Deterministic text smoke: exact POCKETAI_OK
• Deterministic image smoke: exact dominant color red
• Full PocketAI Model Lab suite: 109/109 tests
• Every uploaded variant includes an artifact-manifest.json with exact file sizes and SHA-256 hashes.

The checks above establish artifact integrity, strict runtime loading, basic text generation, and basic image understanding. They are not a broad benchmark or a guarantee of correctness for every prompt, context length, or serving configuration.

License and attribution

These conversions retain the original model's Apache 2.0 license. See LICENSE and the official Qwen model card.

• PocketAI Model Lab: https://github.com/PocketAIHub/pocketai-model-lab