ToPo-ToPo/Ornith-1.5-35B-A3B-mlx-4bit

MLX 4bit conversion of ornith-ai/Ornith-1.5-35B-A3B for Apple Silicon (mlx-vlm), with the vision tower kept.

Since 2026-10-31 this repo uses a mixed-precision recipe: the routed experts are 4bit and every other linear layer is 8bit.

Provenance (self-converted from official weights)

  • Source: ornith-ai/Ornith-1.5-35B-A3B (license: mit)
  • Tool: mlx-vlm 0.7.10 (mlx 0.32.4), mlx_vlm.convert(..., quantize=True, q_bits=4, q_group_size=64, quant_predicate=predicate) with the predicate below
  • Quantization (affine, group size 64):
    • Routed experts (mlp.switch_mlp.{gate,up,down}_proj): 4bit
    • All other linear layers: 8bit. This covers lm_head, embed_tokens, full attention, Gated DeltaNet projections, the shared expert and the routers.
    • Vision tower: not quantized (bf16)
  • Effective: 4.925 bits/weight (weights about 21.6 GB; the previous uniform 4bit was about 20.4 GB)
  • Vision: the source checkpoint is Qwen3_5MoeForConditionalGeneration and contains the vision tower (333 tensors) and the image/video processor configs. This conversion keeps them, so the model accepts image input. The official ornith-ai/Ornith-1.5-35B-A3B-MLX-4bit is a text-only conversion without the vision tower.
  • The MTP (multi-token prediction) weights of the source are not included.
from mlx_vlm.utils import skip_multimodal_module

def predicate(path, module):
    if skip_multimodal_module(path):
        return False
    if ".switch_mlp." in path:
        return {"group_size": 64, "bits": 4}
    return {"group_size": 64, "bits": 8}

Why mixed precision

The previous revision (88f27a0, 2026-09-18) used the default uniform 4bit conversion, with only the routers kept at 8bit. That is the same recipe as the official MLX 4bit. Only 8 of the 256 experts run for each token, while the non-expert layers run for every token. In the uniform 4bit model, Japanese answers sometimes mixed in Simplified Chinese, for example a tool-call preamble written in Chinese. Keeping the non-expert layers at 8bit removed this in our tests. It costs about 6% more weight bits.

Small in-house evaluation, thinking off, 5 samples per case: 34 short Japanese/English probes and 20 tool-calling scenarios, each in Japanese and English. All models were served through local-llm-server (mlx-vlm backend) on the same Apple Silicon Mac.

Metric bf16 (source) previous uniform 4bit this revision
Responses with Simplified Chinese / Hangul (probes) 0.6% 5.9% 0%
Same, Japanese tool calling 0% 4% 0%
Probe accuracy 96.5% 93.0% 95.7%
Japanese tool-calling accuracy 100% 100% 99%
English tool-calling accuracy 100% 100% 100%
Median generation speed 64 tok/s 102 tok/s 91 tok/s

The differences between bf16 and this revision are within one sample. The suite is small and does not cover long agentic tasks.

Validation

  • Image input tested through local-llm-server (mlx-vlm backend). The model correctly named the colors, shapes and positions in a synthetic test image.
  • Tokenization and the chat template give the same token IDs as the source checkpoint on Japanese, English, code, tool-call markup and combining characters.

Usage

from mlx_vlm import load, generate
model, processor = load("ToPo-ToPo/Ornith-1.5-35B-A3B-mlx-4bit")

To use the previous uniform 4bit files, pin the revision 88f27a0ff513bc11fe6b62a5068c3e660639a595.

Downloads last month
174
Safetensors
Model size
35B params
Tensor type
U32
ยท
BF16
ยท
MLX
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for ToPo-ToPo/Ornith-1.5-35B-A3B-mlx-4bit

Quantized
(222)
this model