mlx-community/Sage-A1.5-OptiQ-4bit

Built with mlx-optiq, the MLX-native toolkit to quantize, fine-tune, and serve LLMs locally on Apple Silicon, no PyTorch and no cloud. Try the Lab · All OptiQ quants · Docs

This repository contains a mixed-precision MLX quantization of InternScience/Sage-A1.5 for Apple Silicon. Sage-A1.5 is a 35B mixture-of-experts agentic model from Shanghai AI Laboratory with approximately 3B active parameters per token, designed for long-horizon scientific problem solving, tool use and code. Its architecture is Qwen3.6-35B-A3B: 40 layers, 256 experts, 8 routed per token, with a Qwen3-VL vision tower.

OptiQ quantizes the language tower to mixed 4-bit and 8-bit precision and keeps the vision tower at bf16 in a sidecar file, so a single checkpoint handles both text and images. The checkpoint is 22 GB on disk, down from about 72 GB in bf16.

What it is

Property	Value
Base	InternScience/Sage-A1.5 (Qwen3.6-35B-A3B architecture)
Language tower	per-layer 4/8-bit: 118 tensors at 4-bit, 394 at 8-bit (512 total)
Vision tower	bf16, kept in optiq_vision.safetensors (333 tensors)
MTP head	optiq/mtp.safetensors
On disk	20.6 GB language + 0.9 GB vision + 0.5 GB MTP head
Group size	64
Context	262,144 tokens

Following llama.cpp's naming for mixed quants, the "4bit" label denotes the predominant precision, not the weighted average.

How the bit-widths were chosen

The per-layer bit allocation was not measured on this model. It was copied from mlx-community/Qwen3.6-35B-A3B-OptiQ-4bit, whose allocation came from a KL-divergence sensitivity sweep over a six-domain calibration mix.

This is sound because Sage-A1.5 has the same architecture: same layer count, expert count, hidden size 2048, vocabulary 248,320. Its chat template is Qwen3.6's with one addition, an optional verification-protocol switch. All 512 tensors matched the recipe with none unmatched. An unmatched tensor would silently fall back to flat 4-bit.

Sensitivity reflects a layer's role in the architecture. Fine-tuning can shift it, and that was not measured here. Running optiq convert on the model performs the full sweep.

Sampling

The upstream repository ships no generation_config.json. This repository adds one with the sampling parameters the Sage-A1.5 authors recommend on their model card: temperature 1.0, top_p 0.95, top_k 20, min_p 0, presence_penalty 1.5.

Without presence_penalty, the quant reasoned past 16,384 tokens without answering on two open-ended prompts, a one-sentence explanation and a short coding task. With the recommended values, both prompts finished with correct answers, using about 4,900 and 800 reasoning tokens respectively. optiq serve reads the file automatically. Other runtimes should pass these values explicitly.

What was verified

Run through this quantized model with optiq serve on Apple Silicon:

Input	Output
Capital of France, one word	Paris
17 × 23	391
240 km in 3 hours, average speed	80
"What's the weather in Paris?" with a get_weather tool	get_weather({"city": "Paris"})
An image of an OptiQ release banner	Read the product name, date and listed features correctly
One sentence on why the sky is blue	Shorter blue wavelengths scatter more
An iterative Fibonacci function	Correct, with the base cases handled

Also checked: 512/512 tensors matched the recipe, and the artifact layout, sidecars and metadata follow OptiQ's release contract.

Not run for this model: the six-metric Capability Score. The published scores for the Qwen3.6-35B-A3B quant describe that model, not this one, and are not claimed here.

Use it

pip install mlx-optiq
optiq serve --model mlx-community/Sage-A1.5-OptiQ-4bit --ngram-draft 16

--mtp is not supported on this model yet: the MTP head of the Qwen3.6-35B-A3B family does not load into OptiQ's speculator. N-gram drafting (--ngram-draft 16) is the speed-up to use.

Images through the Python API:

from optiq.runtime.engine import OptiqEngine

eng = OptiqEngine("mlx-community/Sage-A1.5-OptiQ-4bit")
st = eng.generate("What is in this image?", images=["figure.png"], max_tokens=2048)
print(st.text)

Text-only works under stock mlx-lm as well. It globs model*.safetensors and ignores the sidecar, so the same repo loads either way:

from mlx_lm import load, generate

model, tokenizer = load("mlx-community/Sage-A1.5-OptiQ-4bit")
print(generate(model, tokenizer, prompt="Explain KV caching.", max_tokens=2048))

The per-layer bit map is in optiq/metadata.json and in the quantization block of config.json.

Links

• Project website: mlx-optiq.com
• PyPI: pypi.org/project/mlx-optiq
• Base model: InternScience/Sage-A1.5
• Technical report: arXiv 2610.28578