mlx-optiq
← Models

Sage-A1.5-OptiQ-4bit

InternScience/Sage-A1.5, quantized by OptiQ for Apple Silicon; mixture-of-experts, takes images.

visionMoEMTP head
On disk23.5 GB
Memoryabout 28 GB
Parent score80.03 Qwen3.6-35B-A3B-OptiQ-4bit
Imagesyes

This model reuses the per-layer bit allocation of Qwen3.6-35B-A3B-OptiQ-4bit, whose measured Capability Score is 80.03.

The parent's Capability Score breakdown

This model has no Capability Score of its own. The chart shows the six metrics behind its parent Qwen3.6-35B-A3B-OptiQ-4bit's score of 80.03, taken from the parent's model card and compared with uniform 4-bit.

  • Qwen3.6-35B-A3B-OptiQ-4bit
  • Uniform 4-bit
The parent's Capability Score breakdown: Qwen3.6-35B-A3B-OptiQ-4bitMMLU: Qwen3.6-35B-A3B-OptiQ-4bit 83.7, Uniform 4-bit 84.6; GSM8K: Qwen3.6-35B-A3B-OptiQ-4bit 87.9, Uniform 4-bit 89.4; IFEval: Qwen3.6-35B-A3B-OptiQ-4bit 72.6, Uniform 4-bit 73.0; BFCL-V3: Qwen3.6-35B-A3B-OptiQ-4bit 92.5, Uniform 4-bit 90.0; HumanEval: Qwen3.6-35B-A3B-OptiQ-4bit 91.5, Uniform 4-bit 91.5; HashHop: Qwen3.6-35B-A3B-OptiQ-4bit 52.0, Uniform 4-bit 44.0MMLU83.784.6GSM8K87.989.4IFEval72.673.0BFCL-V392.590.0HumanEval91.591.5HashHop52.044.0

Use it

Install

pip install mlx-optiq

Serve it

An OpenAI- and Anthropic-compatible server on this Mac. Every app below can use it.

optiq serve --model mlx-community/Sage-A1.5-OptiQ-4bit

OptiQ Code

The terminal coding agent uses whatever model optiq serve is running.

optiq serve --model mlx-community/Sage-A1.5-OptiQ-4bit
optiq code

OptiQ Lab

optiq lab --model mlx-community/Sage-A1.5-OptiQ-4bit

Claude Code

optiq run claude -m mlx-community/Sage-A1.5-OptiQ-4bit

Codex

optiq run codex -m mlx-community/Sage-A1.5-OptiQ-4bit

OpenCode

optiq run opencode -m mlx-community/Sage-A1.5-OptiQ-4bit

Mistral Vibe

optiq run vibe -m mlx-community/Sage-A1.5-OptiQ-4bit

Hermes Agent

optiq run hermes -m mlx-community/Sage-A1.5-OptiQ-4bit

OpenClaw

optiq run openclaw -m mlx-community/Sage-A1.5-OptiQ-4bit

optiq run starts optiq serve if it is not running, sets the agent up for this session only, and stops the server when the agent exits.

mlx-lm, text only

from mlx_lm import load, generate

model, tokenizer = load("mlx-community/Sage-A1.5-OptiQ-4bit")
print(generate(model, tokenizer, prompt="Hello", max_tokens=256))

Details

FamilyQwen · Qwen docs
ArchitectureQwen3.5 MoE
Base modelInternScience/Sage-A1.5
Size on disk23.5 GB
Memory neededabout 28 GB
Imagesyes, through optiq serve and the Lab
MTP headincluded
Family referenceThis model reuses the per-layer bit allocation of Qwen3.6-35B-A3B-OptiQ-4bit, whose measured Capability Score is 80.03.
Downloads39 in the last month
Updated2026-10-22
Hugging Facemlx-community/Sage-A1.5-OptiQ-4bit