mlx-community/Agents-A1-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

A 4-bit mixed-precision MLX quant of InternScience/Agents-A1, an agentic reasoning model built on the Qwen3.5-35B-A3B Mixture-of-Experts architecture (256 experts, 8 active per token). Sensitive layers are kept at 8-bit and robust ones at 4-bit.

Image input works. The vision tower is kept at bf16 in a sidecar, so this quant takes images as well as text.

65 GB of bf16 weights become 22 GB.

Running it on a 24 GB Mac

At 22 GB this does not fit comfortably in a 24 GB Mac's Metal working set, and loading it resident makes decoding painfully slow. Serve it with SSD expert streaming instead, which reads only the active experts per token:

optiq serve --model mlx-community/Agents-A1-OptiQ-4bit --stream-experts

That brings resident memory down to 4.58 GB. Streaming is the default (auto) in optiq serve, so it engages by itself when a MoE will not fit; the flag above just makes it explicit. On a 32 GB+ Mac the model fits resident and streaming is unnecessary.

Quantization details

Property	Value
Predominant precision	4-bit
Layers at 8-bit (sensitive)	397
Layers at 4-bit (robust)	113
Total quantized layers	510
Achieved bits per weight	4.513
Group size	64
Experts	256 per layer, 8 active per token
Vision tower	bf16, 333 tensors, in optiq/optiq_vision.safetensors
Size on disk	22 GB (language 21 GB, vision sidecar 0.9 GB), from a 65 GB bf16 base

We follow the same naming convention llama.cpp uses for Q4_K_M and similar mixed-precision quants: the "4-bit" label is the predominant precision, not the weighted average.

The base model ships no MTP head, so this quant has no speculative-decoding sidecar.

How the bit-widths were chosen

The per-layer allocation is transferred from mlx-community/Qwen3.5-35B-A3B-OptiQ-4bit, where it was derived by a KL-divergence sensitivity sweep against the bf16 reference on a six-domain calibration mix.

Agents-A1 is built on Qwen/Qwen3.5-35B-A3B and its architecture is unchanged (every field of the text config matches), so all 510 quantizable layers map across exactly and the allocation lands at the same 4.513 bits per weight when recomputed against Agents-A1's own tensors.

These are measured bit-widths, not a static rule-of-thumb recipe. But they were measured on the base model, not on this fine-tune. Fine-tuning shifts weights, so Agents-A1's own per-layer sensitivities could differ somewhat from the base's. Which layers are fragile is mostly a property of the architecture, so the transfer is sound, but it is a transfer and you should know that.

Only the language tower is quantized. The vision tower stays at bf16, which is how every OptiQ VLM ships.

Usage

Text

pip install mlx-optiq
optiq serve --model mlx-community/Agents-A1-OptiQ-4bit --stream-experts

Then point any OpenAI-compatible client at http://127.0.0.1:8080/v1.

The sidecar lives in an optiq/ subfolder, so a stock *.safetensors glob ignores it and mlx-lm sees a clean language model:

from mlx_lm import load, generate

model, tokenizer = load("mlx-community/Agents-A1-OptiQ-4bit")
response = generate(model, tokenizer, prompt="Explain MoE routing.", max_tokens=512)

Note that mlx_lm.load holds the whole model resident, which is slow on a 24 GB Mac. Prefer optiq serve --stream-experts there.

This is a reasoning model: it thinks before answering, so give it enough max_tokens to finish.

Images

Send an image through the OpenAI-compatible endpoint:

import base64, io, requests
from PIL import Image

buf = io.BytesIO(); Image.open("photo.jpg").save(buf, format="PNG")
uri = "data:image/png;base64," + base64.b64encode(buf.getvalue()).decode()

requests.post("http://127.0.0.1:8080/v1/chat/completions", json={
    "model": "a1", "max_tokens": 256,
    "messages": [{"role": "user", "content": [
        {"type": "text", "text": "What is in this image?"},
        {"type": "image_url", "image_url": {"url": uri}}]}]})

Verification

Text, arithmetic reasoning, and image understanding were all exercised on the finished artifact before release, through expert streaming on a 24 GB M4.

The quantization was also checked numerically: dequantizing individual experts out of the artifact and comparing them against the corresponding experts in the bf16 checkpoint gives 0.7% mean relative error on the 8-bit layers and 9.8% on the 4-bit layers, which is what each bit-width should cost.

No task benchmarks were run on this quant; for measured quality numbers on the base architecture, see the Qwen3.5-35B-A3B OptiQ card.

Quantization does not change the behaviour or alignment of the base model. Use it under the same terms as the original.