Neurastheniacc: LLM-based C->WASM compiler you didn't need

A developer released neurastheniacc, an open-source C-to-WASM compiler that uses a local Qwen large language model to handle the C-to-WAT pass, then parses the generated WAT with the `wat` crate and builds it with `wasmi`. The compiler supports only a restricted C subset — one function returning `uint32_t` with exactly two `uint32_t` parameters, no pointers, arrays, loops, recursion or function calls — and caps programs at 32 KiB, 4096 Qwen tokens and 48 nesting levels, with up to 3 retry turns by default when the model emits invalid WAT. On first run neurastheniacc downloads roughly 2.75 GB of Qwen3.5-4B-Q4_K_M weights (the `qwen-low` default) or about 16.5 GB for the 27B `qwen-high` model, running inference locally via llama-cpp-2 with no API keys required.

A groundbreaking C - WASM compiler wielding the unbridled power of LLMs neurastheniacc compiles a restricted subset of C down into WASM. It follows the usual two-pass approach: first from C to readable WAT, then from WAT to WASM bytecode. However, there's a trailblazing advance that distinguishes neurastheniacc from all previous lesser? ∩￢з￢ WASM compilers: the C- WAT pass is handled purely by an LLM. With neurastheniacc , you can finally alchemize money directly into performance, without any need for pesky intermediaries like "compiler expertise" If you're unsatisfied with the defaults' results, try any of the following: - Use a larger Qwen, merely costing more GPU compute - Extend the token limits, reasoning budget, or allowed retries, merely costing more execution time - Build your own stochastically optimizing compiler: run neurastheniacc in a loop 100 times, and pick the best output - Or for the real sickoes, finetune your own C- WAT model and provide a custom GGUF Please wash your hands afterward. plz builb ๑•́ ω •̀๑ cargo build --release --locked --bin neurastheniacc First use downloads ~2.75 GB of Qwen weights, then compiles. ./target/release/neurastheniacc compile examples/mix.c -o mix.wasm Invoke the exported function in a standard Wasm runtime. wasmtime run --invoke mix mix.wasm 3 5 neurastheniacc is fully local. On first run, it downloads a Qwen GGUF and caches it. Inference is then provided during compilation by llama-cpp-2 https://crates.io/crates/llama-cpp-2 . No API keys or calls to inference providers required. During compilation, neurastheniacc validates that your C code is supported, provides your code to Qwen, and asks it nicely to translate it into valid WAT. We then take Qwen's lovingly tensor-crafted WAT just like wire mother used to make and attempt to parse in with wat https://crates.io/crates/wat and build it with wasmi https://crates.io/crates/wasmi . What happens if, Lord forbid, the model should generate invalid WAT? Never fear: a simple harness provides the current state and parser feedback back to Qwen for another attempt, with up to 3 turns allowed by default. neurastheniacc supports this astonishingly flexible program structure: One C function, returning uint32 t , with exactly two uint32 t parameters. An optional include <stdint.h is graciously accepted for your convenience. This luxurious subset of C features is provided, that you might write the programs of your dreams: - Initialized uint32 t locals, assignments, nested blocks and variable shadowing. - + , - , , & , | , ^ , unary ~ , unary + and unary - . - Line and block comments. - << and , with a literal shift value from 0u to 31u . - Braced if / else . All paths must return. If conditions may include: - Unsigned comparisons == , = , < , <= , , = - Integer truthiness - Unsigned comparisons - Decimal and hexadecimal 32-bit integer literals with an explicit u or U suffix. - Returns, including early returns. These extraneous C features aren't supported, and will surely be missed by nobody: - Pointers, allocation, and casting - Arrays - Loops, recursion, and else if - Function calls - Frivolous mathemetics such as division, remainder, increment/decrement, or negative numbers - Globals - Ternaries, or logical operators in general really - Compound assignment, comparison expressions used as values - Macros - Probably a lot of other stuff - Don't even think about trigraphs, IOCCC pervert. Your program must be no larger than 32 KiB and 4096 Qwen tokens. You may not use more than 48 levels of nesting. Select a model: default: qwen-low - qwen-low Qwen3.5-4B-Q4 K M, 4B, ~2.75GB - qwen-high Qwen3.8-27B-UD-Q4 K M.gguf, 27B, ~16.5GB - or whatever local GGUF your little heart desires ./target/release/neurastheniacc compile examples/mix.c --model qwen-low ./target/release/neurastheniacc compile examples/mix.c --model qwen-high ./target/release/neurastheniacc compile examples/mix.c --model /path/to/model.gguf Enable thinking. Only supported with bundled Qwens. default: disabled ./target/release/neurastheniacc compile examples/mix.c \ --qwen-thinking=true Increase the context or per-attempt token budget. ./target/release/neurastheniacc compile examples/mix.c \ --context 16384 --max-tokens 4096 Set limit for compilation attempts. default: 3 ./target/release/neurastheniacc compile examples/mix.c --attempts 5 --verbose Print diagnostics. ./target/release/neurastheniacc compile examples/mix.c --verbose Run on CPU with however many threads. ./target/release/neurastheniacc compile examples/mix.c --gpu-layers 0 --threads 8 Download a Qwen to a custom cache, then compile offline. ./target/release/neurastheniacc fetch --model qwen-low --cache-dir .models ./target/release/neurastheniacc compile examples/mix.c \ --model qwen-low --cache-dir .models --offline Run ./target/release/neurastheniacc compile --help for all options and defaults. The neurastheniacc-bench executable runs a benchmark on the compiler. Eight standard fixtures are provided, alongside as many randomly generated fixtures as your body can handle. Each fixture is run through neurastheniacc with all logs and outputs preserved. For all fixtures that successfully compile to WASM, the resulting executable is proptested against a Clang build of the original C. If we get the same outputs from both neurastheniacc and Clang across a wide set of random inputs, then we consider it a successful compilation °˖✧◝ ⁰▿⁰ ◜✧˖° bliud cargo build --release --locked --bin neurastheniacc-bench Run a benchmark with 32 fixtures, 24 of 'em generated ./target/release/neurastheniacc-bench bench --model qwen-low \ --generated 24 --seed 20261014 --out-dir benchmarks/my-benchmark | Model | Thinking | Pass rate | WAT syntax errors | Wasm validation errors | Semantic mismatches | Token limits | Runtime | |---|---|---|---|---|---|---|---| | Qwen3.5-4B | Off | 71.9% 23/32 | 6 | 1 | 2 | 0 | 2m 13s | | Qwen3.5-4B | On | 21.9% 7/32 | 0 | 0 | 0 | 25 | 3h 41m 48s | | Qwen3.8-27B | Off | 96.9% 31/32 | 0 | 0 | 1 | 0 | 4m 39s | | Qwen3.8-27B | On | 100% 32/32 | 0 | 0 | 0 | 0 | 43m 11s | Each run used Metal on Apple Silicon, with greedy decoding and five retries. Thinking-off runs allowed 8,192 tokens context and 1,536 tokens output budget. Thinking-on runs allowed 16,384 tokens context and 8,192 tokens output budget. 'tis the best kind of joke– an executable joke. Would love to know, please do write my doctor a memo if ya figure it out. You better believe I tried. Didn't work at all well with small local models. The Nix shell provides the full toolchain needed to develop the project: Rust, Cargo, rustfmt, Clippy, rust-analyzer, Clang/libclang, CMake, and Wasmtime. A healthy state of delirium provides your motivation for attempting. neurastheniacc so far has only been tested on macOS with Apple Silicon. It... might work on Linux? Probably not very fast, though. On Mac inference uses Metal, on Linux it uses your poor overheating CPU. You yes, you could probably improve this by contribbing CUDA or something, if you enjoy turning the wheel of samsara. Enter the pinned development environment manually, run checks, and build: duild nix develop Run checks cargo fmt --all -- --check cargo clippy --workspace --all-targets --locked -- -D warnings cargo test --workspace --locked Build the compiler cargo build --workspace --release --locked --bins