Date of scan: 2026-10-24. This distils a survey of comparable/newer projects on GitHub and records the concrete ideas Lume adopts, with sources. “CONFIRMED/UNVERIFIED” tags carry over from the scan.
| Project | What it is | The idea Lume takes |
|---|---|---|
| Bend 2 — github.com/HigherOrderCO/Bend (24.1k★, CONFIRMED) | GPU-parallel language that “blocks AI mistakes via proof”, dependent types, targets C/CUDA/Metal/JS | A project-level LAWS file of compiler-enforced invariants, proof-obligated on every change |
| NanoLang — github.com/jordanhubbard/nanolang (630★, CONFIRMED) | Language “designed to be targeted by coding LLMs”; self-hosting; transpiles to C; verified bytecode (NanoISA); Coq soundness; mandatory shadow test block per function; LSP+DAP+playground |
Mandatory test block per pub function, enforced by the compiler; a verifier-checked bytecode |
| BAML — github.com/boundaryml/baml (9.4k★, CONFIRMED) | DSL for typed LLM functions; built-in tests + eval framework; schema-aligned parsing; colourless concurrency | Eval/mocking framework for extern llm fn + schema-aligned parsing fallback for providers without constrained decoding |
| Mojo — github.com/modular/modular (~30k★, CONFIRMED; compiler open since Aug 2026) | Python-like, CPU+GPU from one source, ownership, in-tree kernel library max/kernels |
Ship a real kernel library with the language as the kernel benchmark |
| Koka — github.com/koka-lang/koka (4.1k★, v3.2.9 2026-09-17, CONFIRMED) | Typed effect handlers (async/exceptions/probabilistic as libraries); Perceus reference counting, GC-free C backend | Implement sample/observe/infer and async as effect handlers, not special cases; Perceus-style reuse analysis |
| Unison — github.com/unisonweb/unison (6.8k★, CONFIRMED) | Definitions identified by AST hash; cached tests, rename-safe edits, abilities (effects) | Content-address definitions (per-AST-hash), not only packages → cached tests and proofs, rename-safe AI edits |
| Austral — github.com/austral/austral (1.6k★, CONFIRMED) | Linear types + capability security; dependencies restricted in which resources they touch; spec-first | Capabilities in package metadata, checked by the package store; keep a small written spec |
| Futhark — github.com/diku-dk/futhark (2.8k★, CONFIRMED) | Pure data-parallel language, guaranteed fusion, GPU+CPU | A parallel-by-construction, fusion-guaranteed array core as the kernel base (tractable under MLIR) |
| Taichi — github.com/taichi-dev/taichi (28.4k★, CONFIRMED) | One source → CUDA/Vulkan/OpenGL/Metal/WASM | Target Vulkan+Metal (not just CUDA) from the start, with a WASM fallback |
| Carbon — github.com/carbon-language/carbon-lang (33.9k★, CONFIRMED, “experimental”) | C++ successor, interop-first migration | C ABI + C-header importer on day one; also a cautionary tale (high stars, no usable release) |
| Zig — github.com/ziglang/zig (43.3k★, CONFIRMED; moved to Codeberg) | Language-native scriptable build system; zig cc cross-compiling C toolchain |
Build system as a first-class language-native tool; compiler as a cross-compiling C toolchain |
| Hylo — github.com/hylo-lang/hylo (1.6k★, CONFIRMED; rewriting as hylo-new) | Mutable value semantics as a simpler alternative to borrow-checker lifetimes | Keep evaluating mutable value semantics to lower the ownership learning curve |
| Redox OS — github.com/redox-os/redox (16.7k★, CONFIRMED, active) | Microkernel OS in Rust | Reference point for lume-os’s capability model (not yet analysed) |
Also noted: MoonBit (human-AI collaboration language; only docs repo found, 2.5k★), SPELL (AI-native dataflow language with JSON source, Zenodo 2025), DSPy (38.8k★, not a language but proves demand for “program, don’t prompt” with prompt optimisers).
pub function (NanoLang). → principle P35, roadmap Stage 1.lume lsp, lume dap, import c).sample/observe/async (Koka) and content-addressed definitions (Unison). → roadmap Stage 3 (handlers) and Stage 2 (content-addressed store already planned; extend to per-definition hashing).model/kernel/llm are macro libraries over a small core (P23), able to lag without blocking the core language.lume agent priority in Stage 1.A scan of GitHub projects that share our name (“Lume”/”Lumi”/”Lumen”/”Luma”) found no feature Lume lacks: on the language-design axis Lume is ahead of every homonym (none has a sound decision procedure, self-hosting with native-ELF emission, or provenance tracking). Kept here as naming and competitive intelligence.
| Project | What it is | Verdict for Lume |
|---|---|---|
| lumeland/lume — github.com/lumeland/lume (2.3k★, active) | Static-site generator for Deno; executable _config.js; ordered process/preprocess passes by extension |
Not a language. Dominates search for “lume”/”lume lang” — the main discoverability risk. Passes/config-as-code are generic; nothing to take. |
| lume/lume — github.com/lume/lume (1.5k★) | 3D/AR/VR toolkit; reactive web-component “behaviors” | Different domain (web UI). Nothing transfers. |
| Lumen — github.com/alliecatowo/lumen (0★, active, Rust) | Self-styled “AI-native language” for LLM agents; effects in signatures, @deterministic, .lm.md markdown source |
Same slogan as us. Effects-in-signature is a subset of our effect rows ([E0200]); @deterministic is subsumed (our / pure already forbids clock/rand/uuid). Uses .lm. |
| rjreeves/lume — github.com/rjreeves/lume (0★) | Small statically-checked script language “optimised for AI generation”; ai-eval.ps1, lume-api.json, versioned system prompt |
Exact homonym, same niche. One genuinely new idea → the LLM-generation eval gate (gap §8 above). |
| Lumi — github.com/meircif/lumi-lang (16★) | Compiles via C; tiered references; per-type integer-overflow policy; built-in tests/mocking; .lm |
Overflow-per-type already covered by our +%/+|/@overflow; tiered refs weaker than our linear ownership ([O0001]). |
| Luma — github.com/Luma-Programming-Language/Luma (89★, self-hosted→C) | #takes_ownership/#returns_ownership annotations instead of lifetimes |
Our linear ownership is stronger; annotation-ownership is a lighter subset. Nothing to take. |
| Lumina — github.com/luminalang/lumina (126★, Rust) | Functional; reflection via const-folding after monomorphisation (a macro alternative) | Same idea as our comptime; already have it. |
| Others — Lumen/Firefly (BEAM→native, archived 3.6k★), laravel/lumen (PHP, archived), rxi/lume (Lua lib), AAsteria/Lumi (λ-calculus) | — | Domain-irrelevant or dead; only naming noise. |
No Spanish-language “Lumi/Lumen” project was found. Lume uses the .lume source extension, so there
is no extension collision with the .lm homonyms.
“AI-first” alone is contested: two 0★ homonyms (alliecatowo/lumen, rjreeves/lume) use the same
slogan, and lumeland/lume owns the search term. Lume therefore leads with what is unique and
verifiable — a sound QF_LIA proof layer, self-hosting with zero C and native-ELF emission,
and provenance tracking — not the slogan. README, GitHub topics and taglines state these first.
Sources: the GitHub repos above, modular.com/blog/mojo-open-source, simonwillison.net/2026/Jan/19/nanolang/ and /2026/Aug/18/mojo-is-now-open-source/, zenodo.org/records/17826541.