Skip to content

Roadmap and status ​

This page separates what exists today from what is planned. Anything under "Planned" is a design in docs/plan-kotlin.md, not something you can use. Sources: the project README and the plan.

Done ​

AreaState
Self-hosting (B3)Done. The native minic-native compiles itself without a JVM or K2; stage 2 and stage 3 are byte-identical, through the C backend and both LLVM backends. CI bootstraps from the latest release.
Own frontend (B2)Done and the default: own parser (0 mismatches against K2's PSI tree on large fuzzed corpora), own type checker (99.99 % agreement with K2 FIR) and lowering to MIR without K2.
Native mini project toolDone. mini new/build/run/test/check, mini.toml, path and git dependencies; a native binary compiled by Mini, about 1 ms to start, shipped in the release archives.
ReleasesPrebuilt archives for linux-x64, linux-arm64 and macos-arm64, plus a C bootstrap seed. See Install.
Memory model checkerDiagnostics M0101 to M0115, M0202, M0204 to M0207 (reference), natively (minic-native --check) and in the browser. M0201 and M0203 are specified but not implemented.
C interopexternal with @CName, CPointer/*Var, memScoped, @CStruct, staticCFunction, @file:CInclude/@file:CLink, @Owned(free=). See C interop.
LLVM backendMIR to LLVM IR, as text or in memory through the LLVM C API bound in Mini; minic-native --run runs programs with the ORC JIT.
Playground and notebook in the browserThe compiler built for WebAssembly runs in your browser (Playground and notebook), and every example in the documentation is a live cell (Learn Mini).
Matrix operator @a @ b calls an operator fun matmul; no type implements it by default. See Mini and Kotlin/JVM.
Performance, first roundRC elision, a size-class allocator and faster arenas, inline array accessors, borrowed array reads, fast integer formatting. Several benchmarks are within 1.0 to 1.3x of C. See Benchmarks.
Language and stdlibclasses, sealed/data/enum/object, interfaces, generics, lambdas, inline/reified, exceptions, extensions, collections. See Mini and Kotlin/JVM.
PlatformsLinux and macOS (CI).

In progress ​

AreaState
Repository without GradleDone: the compiler is a workspace of flat Mini projects, tests run with mini test and scripts/test-e2e.sh, the K2 cross-checks are the separate optional tool tools/k2-oracle (nightly). See docs/no-gradle.md.
LLVM performanceBringing the LLVM backends to at least the speed of the C backend, then making LLVM the default.
Strict codegen and stdlib on the memory model (D2)Strict programs are checked everywhere, and programs that use allocator objects and mini.memory build and run natively; only the samples that use threads remain check-only.

Planned ​

The rows below are not implemented yet.

AreaPlan
Strict values in collectionsList<Box> of owned Strict objects is still M0110. Next steps: release of temporaries per statement in Strict functions (then a temporary listOf(a, b) can borrow its elements), and an owning OwnedList<T> in mini.memory that frees its elements. See the design note in docs/memory-model.md, section 5.4.

Everything in this section does not exist yet.

AreaPlan
Dependencies and packagesDependencies in mini.toml by git URL or path with a lockfile, later precompiled .mlib packages, a build cache and a registry.
ReflectionStatic, read-only metadata for KClass.members, memberProperties, KProperty.get/set, KFunction.call, only for classes that are reflected on, with ownership annotations stored in the metadata. Today only Foo::class and simpleName/qualifiedName exist.
Compiler pluginsA plugin API that generates code at compile time (serializers, DI wiring, route tables).
Libraries outside the stdlibmini-serialization (JSON in the style of kotlinx.serialization), mini-web (a web framework with compile-time DI and routing and one arena per request), tensor libraries on the @ operator. Separate packages, not part of the stdlib.
Coroutines and async IOsuspend lowering, a dispatcher on libuv, mini.io with timers, files, TCP/UDP, DNS. suspend is a compile error today.
LLVM by defaultEmbedded linker, static libuv, .kts scripts and a REPL on the JIT, real exception unwinding.
Performance, second roundA faster HashMap and cached string hashes (wordcount), runtime memory layout, link-time optimization, arenas inside the compiler.
WindowsLinux and macOS first.
Runtime written in KotlinRewrite libminirt piece by piece in Strict mode, and replace libuv step by step.
Cycle collector and weak referencesReference cycles leak today.
Escape analysis and stack placementSpecified (M0201, M0203) but not implemented.
Collections of owned valuesList<@Owned T> and arrays of owned elements; today they are rejected with M0110.
Kotlin librariesSource of Kotlin Multiplatform libraries (commonMain). JVM-only .jar libraries are deliberately not supported.

Plan phases ​

The plan numbers phases P0 to P8 (spike, core backend, runtime and stdlib, memory model, interop, concurrency and IO, libraries, LLVM and performance, self-hosting). The status above is the current view; the original phase table is in docs/plan-kotlin.md, section 10.

Released under the Apache License 2.0.