README.md

Documentation

Every page in this folder, by what you want to do. Back to the README.

Start here

• SETUP.md: what the first run does, starting again and updating, the setup's options, and the fix for each build failure.
• MODELS.md: the four models, how the quants are named, and each one's files, drafters, pictures and options.
• GLM53.md: GLM-5.3 (744B): its files, setup, context and defaults.
• MEMORY.md: which model fits your PC - each model's recommended memory (VRAM + RAM together).
• WINDOWS.md: installing and running Maya on Windows 10/11.

Use it

• SETTINGS.md: the MAYA_* environment variables and config entries that tune the engine.
• RESPONSES.md: the /v1/responses API, its options and limits.
• DOCKER.md: a recipe to build a Docker image of Maya (there is no official image).
• AGENT_BENCH.md: tools/agent_bench.py, a benchmark that loads a running server like coding agents do.
• EXL3-HELPER.md: the experimental helper GPU that runs EXL3 experts the main GPUs miss.

Platforms and hardware

• AMD_MAYA.md: the Linux HIP build for RX 7900 XT / XTX, R9700 / RX 9070 and Strix Halo / Gorgon Halo.
• AMD_RDNA4.md: the HIP installer path for RX 9070 and Radeon AI PRO R9700 (gfx1201).
• AMD_EXPERT_KERNELS.md: which RDNA decode expert kernel the HIP dispatcher picks for each architecture.
• AMD_UPSTREAM.md: an audit of the AMD backend changes shared with another project.
• MI50_VALIDATION.md: the experimental AMD Instinct MI50 (gfx906) source build.
• METAL.md: the experimental Apple Silicon (Metal) build.
• PORT-MINIMAX-METAL.md: notes on porting MiniMax-M3.1-Flash to Metal.
• DGX_SPARK.md: Maya on an NVIDIA DGX Spark (GB10) and Grace, where memory is shared with the system.
• DGX_SPARK_BUILD.md: what the aarch64 port for the DGX Spark changed and where each part comes from.
• IBM_AC922.md: Maya on the IBM Power System AC922 (POWER9, ppc64le, six V100s).
• INTEL_ARC.md: a community report of Maya's DeepSeek GGUFs on llama.cpp's SYCL backend; outside Maya's engine.

Engineering notes

How the engine works and how changes were measured. You do not need these to run Maya.

• GLM5-FLASH.md: the study of what it takes to run GLM-5.3-Flash on the engine, component by component.
• QUANT-PLAN.md: the plan for Maya's first GLM-5.3-Flash quants.
• GLM_MMQ_CORRECTNESS.md: the incoherent GLM prefill replies on Windows with CUDA 13.2.
• GLM-KDA-GRAPHS.md: MAYA_GLM_KDA_GRAPH=1, experimental graph capture of GLM's recurrent layers.
• deepseek-prompt-format.md: DeepSeek-V4.1's prompt encoding and the vendored serve/ds41_encoding.py.
• ds4-draft-timing.md: the DSpark drafter timing fields in DeepSeek STAT lines.
• ds4-skip-unpicked.md: DeepSeek prefill transfers of only the experts a chunk picks (DS4_SKIP_UNPICKED).
• DS4_PROFILE_REFRESH.md: refreshing DeepSeek's startup expert placement, measured on a three-GPU AMD PC.
• EXL3_HIP.md: EXL3 prefill on HIP.
• EXL3_HIP_SCALAR_ROWS.md: configurable scalar EXL3 prompt tiles on RDNA4.
• EXL3_RDNA4.md: matrix-instruction EXL3 prefill on RDNA4 (EXL3_WMMA_GFX12).
• HIP_PEER_PREFILL.md: how startup picks the HIP prompt transport between two GPUs.