Support Ternary Bonsai 2 27B (MLX 2-bit) with reasoning and vision #377

Request

Please add support for prism-ml/Ternary-Bonsai-2-27B-mlx-2bit in LM Studio's MLX engine, including reasoning (thinking mode), vision, and tool use.

The model loads, but the reasoning toggle / thinking-mode parsing is not available for it in LM Studio.

Context

Bonsai 2 27B is Prism ML's ternary (2-bit) compression of Qwen3.8-27B (~1.72 bits/weight, ~98.2% of FP16 quality). The pieces seem to already be in place:

• The first-generation prism-ml/Ternary-Bonsai-27B-mlx-2bit is supported by LM Studio (listed on https://lmstudio.ai/models/prism-ml/bonsai-27b with Reasoning capability) — see also Please support for Ternary Bonsai 27b with vision an reasoning #347.
• The base model qwen/qwen3.8-27b is supported, including reasoning.

What's missing for Bonsai 2 specifically:

• Its chat template (Qwen3.8-based, with reasoning_effort xhigh/medium and enable_thinking) is not recognized, so no reasoning controls are exposed and thinking blocks are not parsed.
• Bonsai 2 weights apply additional transforms (blockwise Hadamard rotation, sign flips, model graph changes). The MLX pack ships as plain safetensors and reportedly runs on stock mlx-lm, while the GGUF formats still require Prism ML's llama.cpp fork (upstream integration pending: https://github.com/PrismML-Eng/Bonsai-demo/blob/main/BACKEND-SUPPORT.md).
• There is no Bonsai 2 entry on lmstudio.ai yet (https://lmstudio.ai/models/prism-ml/bonsai-2-27b returns 404).

Why it matters

A 27B-class reasoning model in ~8.6 GB on Apple Silicon would be a huge win for local inference in LM Studio, just like the first-gen Bonsai was.

Thank you!