Platform Support

Platform Support

Last updated: 2026-08-04.

The authoritative status for the 20 ARM, Intel and portable-engine recommendations is docs/PLATFORM-SUPPORT.json. This page summarizes that machine-checked source. QEMU results are correctness and ABI evidence only.

Recommendation Status

#	Recommendation	Status
1	Restore green CI	ci-tested
2	ARM SMP worker dispatch	qemu-tested
3	ARM FP and NEON context preservation	qemu-tested
4	Native macOS and Linux engine executable	hosted-tested
5	Physical Apple NEON evidence	physical-gate
6	Generic ARM server scope	scope-defined
7	SVE and SVE2	interface-only
8	x86 AP startup and worker participation	qemu-tested
9	x86 ring 3, syscalls and user threads	partial
10	x86 PCI storage, network, NVMe and interrupts	partial
11	x86 full platform services	pending
12	MADT, SRAT, SLIT and HMAT	parser-tested
13	XSAVE and XRSTOR state management	qemu-tested
14	Physical Intel and Xeon evidence	physical-gate
15	Inference engine service ownership boundary	hosted-tested
16	Immutable 64-bit model mappings	qemu-tested
17	Async direct model range I/O	hosted-tested
18	Lifecycle-safe sessions	interface-tested
19	Documentation reconciliation	synchronized
20	GitHub milestone reconciliation	synchronized

The exact evidence for each row remains in the authoritative JSON source.

Implemented QEMU And Hosted Foundations

• AArch64 starts runtime-discovered CPUs, dispatches joinable workers, and preserves q0-q31, FPCR and FPSR across live timer interrupts.
• The native macOS/Linux engine executable exposes probe/inspect and a fail-closed serve command through caller-owned model/session registries.
• x86 starts all MADT-discovered application processors through an OS-owned trampoline and dispatches deterministic IPI work without a project-level CPU count limit.
• x86 validates GDT/TSS ring-3 syscall entry, runtime-sized XSAVE/XRSTOR with FXSAVE/FXRSTOR fallback, ACPI MADT/SRAT/SLIT/HMAT parsing, modern VirtIO block DMA, MSI-X completion, and VirtIO network TX.
• Production model registration retains immutable readers or no-copy 64-bit mappings. Direct aligned range I/O and lifecycle-safe 64-bit session metadata pass hosted tests. The copied model arena remains fixture-only.
• Aggregate CI pins upstream QEMU commit 6ce361b02c825b4a12a9684c47342859ee967cb2 for the translated SMMUv3 gate, verifies its test-only iommu-testdev, and fails rather than skipping when that device is unavailable.

Deliberate Open Boundaries

• Full ARM-service parity on x86 remains open: the complete EL0 process/thread ABI, receive networking, SSH/control, mounted filesystems, x86 NVMe operation, security, AI Cell and telemetry are not linked as one x86 service image.
• SVE/SVE2, Metal, AVX-512/VNNI and AMX are capability/roadmap entries, not executing production backends.
• Typed model state, ragged batching, exact speculation, tokenizer parity, logits parity and real-model decode are not implemented.
• Physical Apple, Intel Desktop and Xeon validation cannot be established by QEMU and remains a separate required gate.

The repository tracker and hardware-readiness document retain these open items; an interface, parser, or QEMU canary is not a production-support claim.

The remaining implementation blocker for full x86 service parity is tracked in issue #18. Physical Apple, ARM, Intel and Xeon validation is tracked separately in issue #19.