Qwen3.8-27B-4bit-MTP-MLX

Run it on your Mac

curl -fsSL https://rapidmlx.com/install.sh | bash
rapid-mlx serve qwen3.8-27b-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 · 32 GB unified memory recommended for the 16.3 GB target and MTP weights.

Qwen3.8-27B 4-bit MLX + MTP

An official Rapid-MLX, self-contained MLX package of Qwen3.8-27B. It includes both the 4-bit target weights and the matching native Multi-Token Prediction (MTP) drafter in model-mtp.safetensors.

Users download one repository; the target and acceleration weights stay paired. The MTP weights were derived from the same upstream Qwen checkpoint as the target. Upstream model license and limitations continue to apply.

Rapid-MLX

Use a Rapid-MLX release that supports Qwen3.8 native MTP. Rapid can load the target and use the colocated model-mtp.safetensors as its MTP sidecar. MTP is speculative decoding: accepted output remains identical to greedy autoregressive decoding. Runtime should fall back to autoregressive decoding when MTP is unavailable or is not beneficial on the current Mac.

rapidmlx serve rapid-mlx/Qwen3.8-27B-4bit-MTP-MLX \
  --speculative-config '{"method":"mtp","model":"rapid-mlx/Qwen3.8-27B-4bit-MTP-MLX"}'

Validation

On an Apple-silicon Studio test system, a paired greedy smoke test produced the same emitted-token hash with and without MTP. One measured prompt ran at 39.66 tok/s autoregressive and 46.73 tok/s with Rapid MTP (1.18x). Performance varies by prompt and hardware; this is a compatibility result, not a universal speed claim.

Contents

• model-*.safetensors: Qwen3.8-27B MLX 4-bit target weights
• model-mtp.safetensors: matching MLX 4-bit native MTP drafter
• tokenizer, processor, generation, and model configuration files

The target is based on mlx-community/Qwen3.8-27B-4bit; the MTP drafter is based on mlx-community/Qwen3.8-27B-MTP-4bit. Both originate from Qwen/Qwen3.8-27B.