This variant does not include the source model's MTP tensors. The MTP config fields are normalized to zero so it loads as the plain text-generation model.

The tokenizer includes the qwen3_coder tool parser setting and a patched chat template for OpenAI-style tool calls. The template was adjusted for Swival-style coding-agent work, including array arguments for command tools and read-before-delete/edit behavior.

Local Validation

Validated on June 12, 2026 with oMLX:

Static artifact check passed: oQ8 quantization, no MTP tensors, matching chat_template.jinja and embedded tokenizer_config.json template.
Direct OpenAI-compatible tool-call smoke passed for scalar string arguments and argv-array arguments.
Swival tool-calling gate passed 5/5 cases with all tools enabled: edit by line number, read/edit/test, batch read/write, read/delete, and argv-form command execution.

Use it through oMLX with model id:

Qwopus3.6-27B-Coder-8bit.mlx

jedisct1/Qwopus3.6-27B-Coder-8bit.mlx

Qwopus3.6-27B-Coder 8-bit MLX

Local Validation

jedisct1/
Qwopus3.6-27B-Coder-8bit.mlx