Language Profiles
Three compile-target profiles. Same source, three artifacts, one stable spec. Pick the profile by what you're shipping — production ML, governance kernels, or bare-metal embedded — and mindc emits a binary sized and instrumented for that target.
mind build --profile=<default|systems|embedded>Flag shipped in mindc 0.2.6; cross-profile linker refusal still on the roadmap; pure-MIND stdlib (RFC 0005) auto-bundled since 0.4.2. Current release line: v0.10.x.
default
frontend latency (parse + typecheck + IR)
Full tensor stdlib, Q16.16 throughout, heap allocator on. The production ML profile — designed for commercial GPU / accelerator backends (roadmap).
- Full tensor stdlib (math, linalg, autodiff)
- Q16.16 fixed-point throughout
- Heap allocator (dynamic shapes)
- —Commercial GPU / NPU / TPU backends (roadmap)
systems
frontend latency
Stripped tensor stdlib + Q16.16 + governance primitives + heap. The control-plane and DIFC-governance profile.
- Stripped tensor stdlib (no autodiff, no linalg)
- Q16.16 + governance kernel
- Heap allocator on
- Invariant system + audit hash primitives
embedded
frontend latency
Stripped tensor stdlib, no heap, Q16.16, governance. Cortex-M / RISC-V / WASM bare-metal — MIND for environments where malloc is not available.
- Stripped tensor stdlib
- Q16.16 + governance kernel
- No heap (all stack / static)
- Cortex-M / RISC-V / WASM targets
At a glance
| Capability | default | systems | embedded |
|---|---|---|---|
| Frontend latency | 1.8–15.5 µs | 0.8–3 µs | 0.4–1 µs |
| Tensor stdlib | Full | Stripped | Stripped |
| Q16.16 fixed-point | |||
| Heap allocator | — | ||
| Governance kernel | — | ||
| Commercial GPU / accelerator backends | roadmap | — | — |
| Target environments | Server, multi-node (GPU via commercial runtime) | Control plane, governance, agents | Cortex-M, RISC-V, WASM |
Read the spec
Profile selection is formalised in mind-spec v1.0 Future Extensions § Language Profiles; implementation lands in mindc Phase 10.6.