PocketAiHub/PocketAI-Ornith-1.0-35B-MLX

PocketAI Ornith-1.0-35B MLX

Official-source MLX releases of deepreinforce-ai/Ornith-1.0-35B, converted and validated by PocketAI Model Lab. Ornith is a reasoning model for agentic coding, post-trained from Qwen3.5-35B-A3B. It is a multimodal mixture-of-experts model with 35B total parameters, approximately 3B active parameters, 256 experts, and 8 routed experts per token.

Variants

Variant	Folder	Stored size	Precision layout
MLX 4-bit	4bit/	20,429,175,098 bytes (19.03 GiB)	432 language/MoE modules Q4; 80 router/shared-expert gates Q8; vision tower BF16
MLX 8-bit	8bit/	37,748,374,364 bytes (35.16 GiB)	512 language/MoE modules Q8; vision tower BF16
MLX BF16	bf16/	70,241,342,643 bytes (65.42 GiB)	All 1,066 stored tensors BF16

All quantized modules use MLX affine quantization with group size 64. The 4-bit release intentionally retains 80 router and shared-expert gate modules at 8-bit precision; this is the converter's quality-preserving MoE layout. The converter reports effective stored precisions of 4.649 and 8.596 bits per weight for the 4-bit and 8-bit releases respectively.

Creative coding showcase

Each variant received the same prompt to create a colorful, single-file HTML voxel pagoda garden. The synchronized comparison below uses the same camera path for all three versions.

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	95.98 tok/s	21.20 GB	11,504	120.44 s
MLX 8-bit	64.96 tok/s	38.92 GB	13,128	203.84 s
MLX BF16	47.72 tok/s	70.59 GB	12,491	268.59 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 each variant emitted the same tokens.

The 8-bit HTML passed its original browser run. The 4-bit and BF16 generations contained small JavaScript defects; the video uses minimally repaired copies while preserving the original benchmark outputs. Exact settings, raw measurements, prompt text, and repair status are recorded in 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-Ornith-1.0-35B-MLX"
variant = "4bit"  # "4bit", "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,
    "Write a short Python function that checks whether an integer is prime.",
    num_images=0,
    enable_thinking=True,
)
result = generate(
    model,
    processor,
    prompt,
    max_tokens=512,
    temperature=0.6,
    top_p=0.95,
    enable_thinking=True,
)
print(result.text)

Ornith is a reasoning model. With thinking enabled, output can contain a <think>...</think> block before the final response. Thinking can be disabled by passing enable_thinking=False to both prompt construction and generation.

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

Reproducibility and validation

• Official source: deepreinforce-ai/Ornith-1.0-35B
• Pinned source revision: 5df2ed3f675c7beaa490328cc70bb573b65fb660
• Converter: mlx-vlm==0.6.8
• Base dtype: BF16
• Quantization: MLX affine, group size 64
• Preserved MoE layout: 256 experts, 8 routed experts per token
• Deterministic text smoke: exact POCKETAI_OK for every variant
• Deterministic image smoke: exact dominant color red for every variant
• Peak MLX memory: 21.14 GB (4-bit), 38.76 GB (8-bit), 70.44 GB (BF16)
• Full PocketAI Model Lab suite: 101/101 tests
• Every uploaded variant includes an artifact-manifest.json with exact file sizes and SHA-256 hashes.

The standalone MTP draft weights from the source checkpoint are not included; these are standard MLX generation artifacts, not a speculative-decoding MTP bundle.

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, tool-use flow, or serving configuration.

License and attribution

The upstream model card declares the source model under the MIT license. See LICENSE and the official Ornith model card.

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