mlx-community/ACE-Step-1.5-planner-1.7B-4bit

ACE-Step 1.5 planner 1.7B (the 5 Hz language model) — MLX, 4-bit (affine, group 64). Converted from ACE-Step/Ace-Step1.5 @ 19671f4 by the converter of xocialize/mlx-ace-step-swift (Tools/convert/convert.py, mlx 0.32.4), a full port of ace-step/ACE-Step-1.5 @ ca1e85f to Swift-MLX.

What is in the bundle

The ACE-Step-1.5-planner-1.7B-bf16 bundle with the 28 layers' 196 Linears quantised — affine, group size 64, 4 bits, mx.quantize (the same core op MLX Swift's QuantizedLinear.quantize calls, so every tensor is bit-equal to what the package produces when it quantises the bf16 planner at load to 4 bits — verified tensor for tensor, 702 / 702) — stored as weight (packed uint32) + scales + biases per Linear in mlx-lm's layout, with "quantization": {"group_size": 64, "bits": 4, "mode": "affine"} in config.json; the 217 204 × 2048 tied embedding / head and the norms stay bf16. 1.68 GB on the Hub (model.safetensors 1 682 718 701 bytes) against the 8-bit bundle's 2.39 GB and the bf16 bundle's 3.71 GB — the smallest download. The package's .int4 tier (offered since v0.3.0; .int8 stays the default): on the port's 12-song gate the 4-bit planner read PER 0.212 against the 8-bit planner's 0.108 and the bf16 planner's 0.120–0.197 across draws (its MEASUREMENTS.md §V2b), and the operator's listening pass over the takes cleared it ("lyrics and quality were great with everything I checked"). The Qwen3 base is Apache-2.0 (Qwen); the fine-tune is ACE-Step's (MIT).

Licence

MIT. The Hugging Face repository ACE-Step/Ace-Step1.5 ships no LICENSE file; its model-card tag and README state MIT and the GitHub LICENSE (Copyright (c) 2026 ACEStep) is the controlling text. The Qwen3 base model is Apache-2.0.

The full notice travels with the bundle as NOTICE; LICENSE is the MIT text. Nothing in the weights is modified beyond the layout conversion (key prefix / conv layout / weight-norm folding) — the converter Tools/convert/convert.py in xocialize/mlx-ace-step-swift is gated bit-equal against upstream's own dit_convert.py / vae_convert.py (its MEASUREMENTS.md §V1).

Use

Every tier of the Swift package loads four bundles — the turbo DiT, one planner, the text encoder and the VAE — and this bundle is one of them. The planner choices: ACE-Step-1.5-planner-1.7B-8bit for .int8 (the default), -4bit for .int4, -bf16 for .bf16 / .fp32, and ACE-Step-1.5-planner-0.6B-bf16 for the lite planner on any tier. The package, xocialize/mlx-ace-step-swift, runs ACE-Step 1.5 on Apple Silicon through MLX Swift and plugs into MLXEngine on the music capability (caption + lyrics → a 48 kHz stereo song; reference-audio timbre; cover / repaint / extend of a source clip; LRC timestamps). The package's ACEStepConfiguration names these repos and the engine materialises them; the parity gates against the upstream oracle (fp32, TF32 off) are recorded in the package's MEASUREMENTS.md.

let engine = MLXServeEngine()
let id = try await engine.register(ACEStepPackage.registration, configuration: ACEStepConfiguration(quant: .int4))
try await engine.prepare(.music, package: id)
let song = try await engine.run(MusicRequest(caption: "…", lyrics: "[Verse]\n…", durationSeconds: 120, seed: 7), package: id)

The package loads this bundle as already quantised for its .int4 tier: the loader builds the QuantizedLinears from the bundle's quantization config before the weights land (nothing is re-quantised at load).

Upstream: ACE-Step/Ace-Step1.5 @ 19671f4 (weights), ace-step/ACE-Step-1.5 @ ca1e85f (code) — ACE Studio + StepFun, tech report.