[smalinin]

What this changes

Adds interactive setup for downloaded GGUF models, including directory/file browsing, editable model/drafter/vision selection, and Esc navigation.

• Supports linking unlisted quants without copying weights or checking published sizes/SHA-256.
• Adds DeepSeek and MiniMax GGUF DSpark/vision companions and distinct model IDs per quant.
• Extends Qwen quantization support and fixes PLE precision and split-shard expert loading.
• Supports both GLM vision formats, Q5_K DFlash2 selectors, and 256K/512K GLM context options.

Answers

Correctness fixes change outputs for affected Qwen GGUF layouts that previously produced corrupted text.

Tested on

Linux/CUDA with RTX 4090/4090D and RTX 3090. Validation includes setup/packer regression tests, CPU/GPU numerical checks, and real-model generation, drafting, and vision tests.

Checks

• Builds on Linux/CUDA.
• Relevant automated tests pass.
• Local model setup documentation updated.

Tested models

• MiniMax-M3.1-Flash - Q2_K
• [MiniMax-M3.1-Flash - Q3_K ]
• GLM-5.3 - UD-IQ3_XXS
• GLM-5.3-Flash — UD-Q5_K_XL
• GLM-5.3-Flash-Uncensored — Q4_K_M
• DeepSeek-V4.1-Flash-1017 — IQ2_XXS
• [DeepSeek-V4.1-Flash-1017 — MXFP4]
• [DeepSeek-V4.1-Flash-1017 — Q3_K]
• Qwen3.8-Flash-Next-Uncensored — Q8_0
• Qwen3.8-Flash-Next — Q8_0