Run the paired model on your Mac

This repository is a sidecar, not a standalone model. The command serves its paired target and uses this sidecar.

curl -fsSL https://rapidmlx.com/install.sh | bash
rapid-mlx serve glm5.3-flash-4bit
curl -s http://localhost:8000/v1/chat/completions -H 'Content-Type: application/json' -d '{"model":"default","messages":[{"role":"user","content":"Say hello"}]}'

Use it from OpenCode / Claude Code / Codex: rapid-mlx agents <name> --setup · Agent setup

Download the Desktop app · 192 GB unified memory for the paired target plus this 4.2 GB drafter. This repo alone cannot answer prompts.

GLM-5.3-Flash MLX Q4 MTP drafter

This repository contains the standalone, target-matched MTP drafter used by Rapid-MLX for Vontra/GLM-5.3-Flash-MLX-4bit-MTP.

It is a sidecar head, not a standalone language model. The target and drafter must be loaded together by a runtime that implements GLM-5.3's cache-owned MTP transaction.

Immutable pairing

Item	Value
Target	`Vontra/GLM-5.3-Flash-MLX-4bit-MTP`
Qualified target revision	`76add2a341a1cd90ad0e86bb69839ea9c35827c6`
Drafter architecture	`glm5_next_mtp`
Block size	2 (one proposed token plus target bonus)
Quantization	4-bit affine, group size 64
Tensor count	55
Weight size	4,183,323,401 bytes
Weight SHA-256	`369cf9c0f9cdf3ae5f1b9f72d3db8e65ad3026b8e416b8de5618e9117d3f00ca`

Do not pair this drafter with another target revision or quantization. Rapid-MLX checks the immutable target/drafter revisions, architecture, and block size before loading.

Rapid-MLX qualification

Measured on an Apple M3 Ultra with 256 GB unified memory, single-row greedy decoding, temperature 0, and the target revision above:

Six real tasks covered coding, knowledge, math, instruction following, creative writing, and long-context document retrieval.
Two consecutive runs passed 12/12 tasks.
Reasoning and final responses were byte-identical to the autoregressive control on all six tasks.
Per-task speedups were 1.405x, 1.377x, 1.306x, 1.390x, 1.285x, and 1.106x.
Paired median throughput improved 1.341x (+34.1%).

Sampling, penalties, logprobs, and batch decoding are not qualified for this artifact. Unsupported request shapes should use ordinary autoregressive decoding.

Provenance

The drafter was split from the native MTP layer preserved in the qualified target checkpoint and converted to the target's affine Q4 representation. The target is itself a quantization of zai-org/GLM-5.3-Flash. See the target model card for upstream provenance, architecture, intended use, and limitations.

License