Linux x86-64 / AMD Strix Halo gfx1151 is the production target. Build with nix build; Python/reference toolchains are development-only. Model weights are downloaded separately and are not part of the runtime package.

Model	Weights / supported inputs	Guide
DeepSeek V4 Flash	Flash 0731 mixed IQ2/Q2/Q8 GGUF; text	Usage and modes
EmbeddingGemma 2	BF16 safetensors; text/image to embeddings	Usage
MiniMax H3 FL2VA	Pinned safetensors; text to audiovisual output	Presets and usage
Qwen-Image-2.1	BF16 safetensors; generation and image editing	Usage
Qwen3-ASR 1.7B	BF16 safetensors; audio to text	Usage
Qwen3-TTS 12Hz 1.7B	BF16 safetensors; text/reference audio to speech	Voice modes
Qwen3.8 27B	Q4/Q8 GGUF; text/images	Usage and modes
Qwen3.8 Flash-Next	Sharded Q4 GGUF; text/images	Usage and modes

Each model folder contains:

README.md: model card, acquisition, usage and modes.
BENCHMARKS.md: current retained measurements, scope/date and TODO cells.
QUALITY.md: measured quality, reference/scope, replay guarantees and unresolved gaps.
EXPERIMENTS.md: short retained/rejected decisions.
artifacts/: only useful machine-readable results/calibration, when present.

Keep executable fixtures with their tests. New logit/trace dumps and local profiles belong in ignored top-level artifacts/; commit only independent reference evidence needed by a maintained check. Old experiment records remain in Git history. See benchmark methodology.

Quality reports distinguish agreement with an independent model implementation from consistency between Gufo execution modes. Greedy speculative decoding should reproduce AR; sampled speculation can use different random draws. Seeded replay is checked within the same execution configuration. For numerical comparisons, lower KL divergence, total variation, relative L2 and word-error rate are better; higher cosine similarity and image PSNR are better. These bounded checks are not broad capability or perceptual scores.