README.md

Model recipes

Read a model recipe when a checkpoint needs extra files, model-specific flags, or when it has a known limitation. Read the quickstart to run your first model, and the usage guide for the shared CLI, server, C ABI, and source-tree workflows.

Each page tells you the same four things in the same order: what the model is, which checkpoint was used, the exact command, and what has not been measured.

Text generation

Model	What it is	Read the page for
Qwen3.8 2.4T	A 2.4-trillion-parameter mixture of experts, 370 GiB at UD-Q1_0	Serving a checkpoint three times larger than the machine's memory, at 11.05 s/token
Qwen3.8 27B	A 27B dense model	Which quantized arms run, and why block-wise FP8 is CPU-only today
Qwen3.6	The Qwen3.6 dense and MoE family	Which lm_head forms load, and the merged FP8 in_proj_qkvz GEMM
Qwen3.5	The Qwen3.5 Gated DeltaNet family	The output_gate_type key, and one load refusal that is about this code
Qwen3-Next	The Qwen3-Next Gated DeltaNet family	The output_gate_type key and its refusals
Nemotron 3.5 Lightning	A 30B-A3B hybrid of attention, Mamba2, and MoE	Which arms run on the device, which run on the host, and what that costs per token
Muse Glimmer	A 30B multimodal model, text with image and video input	Running the text tower from a 17 GB GGUF, and how narrow the verified surface is
Gemma 4	The Gemma 4 family	The ROCm RDNA4 dual-GPU FP8 recipe

Speech, music, and video

Model	What it is	Read the page for
IndexTTS 2.5	Text to speech, 22.05 kHz mono	Converting the checkpoint and starting a speech-only server
MiniMax-Music3	Music generation	The checkpoint set and the render command
MiniMax-H3	Video with audio	The five-file checkpoint set and the render workflow
LTX 2.5	Video generation	The checkpoint set, the render command, and what a render costs

Speculative decoding

Model	What it is	Read the page for
DSpark	Draft models for speculative decoding	Which draft layouts run, and which are refused by name

What these pages do not carry

A model page records what was measured and what was not. It does not carry the design rationale for a default or the history of a fixed defect. Those live in the row's spec under .agents/specs/, and each page links to its own.

You want	Read
What is proven and what is not	Project status
The complete measurement record	docs/BENCHMARKS.md
Which features, backends, and quantizations exist	docs/FEATURES.md
Every server flag and endpoint	server reference
Every environment knob	docs/ENVIRONMENT.md