Runtime matrix

Supported bundled-runtime platforms

Host	Architecture	Accelerator	Runtime location	Status
macOS	Apple Silicon (`arm64`)	Metal / unified memory	Host-native vLLM Metal	Supported
Linux	x86_64	NVIDIA CUDA	Docker Compose `runtime` service	Supported
Windows	—	—	—	Not supported
macOS Intel	`x86_64`	—	—	Not supported
Linux without NVIDIA	x86_64	CPU or other GPU	External provider only	No bundled runtime

The API stack can still use a separately operated provider when the bundled runtime is unavailable. Start with parsehawk start -x runtime.

Platform	Verified	Recommended minimum
macOS	M3 Pro with 18 GB and 36 GB unified memory	16 GB unified memory
Linux	NVIDIA L4 with 24 GB VRAM	16 GB VRAM

32 GB unified memory or 24 GB VRAM gives more room for larger contexts.

Setting	Default
Model	`numind/NuExtract3-W4A16`
Maximum generated tokens	`2048`
Base maximum model length	`8192` before platform-profile adjustment
Base maximum sequences	`1` before platform-profile adjustment
Linux vLLM package/image	`vllm==0.23.0` / `vllm/vllm-openai:v0.23.0`
PDF pages	`25`
PDF render DPI	`170`
Model request timeout	`600` seconds

ParseHawk applies platform- and memory-tier runtime profiles at startup. The effective values can therefore differ from the base settings. Run:

parsehawk runtime info --json
parsehawk config list --json