Clojure, without the JVM.

cljrs is a Clojure dialect implemented from scratch in Rust. The reader, evaluator, macros, and persistent collections run as a tree walker. Type-hinted functions compile to native code through MLIR and LLVM. GPU kernels compile to WGSL and run on Metal, Vulkan, DX12, or WebGPU.

One language, three compilers

;; Tree walker. Fully dynamic, lowest ceremony.
(defn avg [xs] (/ (reduce + xs) (count xs)))

;; Native. Same syntax, type-hinted, JIT-compiled to machine code.
(defn-native fib ^i64 [^i64 n]
  (if (< n 2) n (+ (fib (- n 1)) (fib (- n 2)))))

;; GPU. Same syntax, compiles to a compute shader.
(defn-gpu doubled ^f32 [^i32 i ^f32 v]
  (* v 2.0))

Run it

Status

• Tree walker, macros, persistent collections, destructuring, atoms, try/catch. clojure.core 405/662, plus 100% of clojure.string, clojure.set, clojure.walk, clojure.edn. See Coverage.
• Native code path: 2.4x to 13x faster than Clojure on JVM on fib, loop_sum, cond_chain (as of 2026-04-16). See Benchmarks.
• GPU path: cljrs kernels run natively on Metal/Vulkan/DX12 and in the browser on WebGPU.
• Rust-crate interop via opaque handles — rapier2d/3d bindings ship as Physics.
• 21 CPU demos, 7 GPU demos, all live-coded. See Demos.
• Live matmul benchmark vs numpy/JAX/PyTorch references — Matmul bench.
• New: edit Conway's Game of Life rules live without losing the simulation.