Qwen3.8-27B-mlx-bf16

ToPo-ToPo/Qwen3.8-27B-mlx-bf16

MLX bf16 conversion of Qwen/Qwen3.8-27B for Apple Silicon (mlx-vlm).

Provenance (self-converted from official weights)

• Source: Qwen/Qwen3.8-27B (license: apache-2.0)
• Tool: mlx-vlm 0.6.9 — mlx_vlm.convert --hf-path Qwen/Qwen3.8-27B --mlx-path . --dtype bfloat16
• Effective: bfloat16 (unquantized)
• Architecture: Qwen3_5ForConditionalGeneration (model_type: qwen3_5), 64 layers with 3:1 linear/full attention, vision tower included (333 vision tensors).

Validation

• Text and image input verified with mlx_vlm.generate.
• OpenAI-compatible gateway (local-llm-server): text, image (data URI) and native tool calling (finish_reason: tool_calls) all verified.
• End-to-end agent runs (MCP tools, write → read → report) completed without errors.

Throughput and memory use depend on your hardware, so measure them on your own machine.

Note on MTP

The official checkpoint declares mtp_num_hidden_layers: 1, but the mlx-vlm qwen3_5 conversion does not carry the model.mtp.* weights. A drafter split from the official bf16 checkpoint is published separately as ToPo-ToPo/Qwen3.8-27B-MTP-bf16 (a 4bit variant is also available).

Usage

from mlx_vlm import load, generate
model, processor = load("ToPo-ToPo/Qwen3.8-27B-mlx-bf16")