NURL registrynurl-lang.org →45 package(s) · get a publish token →
0.5.3 — Streaming anomaly-detection service in pure NURL — dynamically created, automatically trained models over Isolation Forests. Where the `iforest` package is the 0.6.0 — Audio for speech models, in pure NURL: read WAV, resample it honestly, and compute the log-mel spectrogram. The WAV reader treats the file as untrusted input — 0.1.2 — Reproduce NURL's performance claims on your own machine — a runnable benchmark suite, not a number in a README. `benchmark` measures pure-NURL throughput for th0.1.0 — Content-addressed store over BLAKE3 — the keystone of a verifiable build/distribution chain: bytes in, their BLAKE3-256 hex out, and from then on the hash IS th0.1.1 — Terminal data-visualisation — sparklines, bar charts, histograms, and line plots, as a CLI and a reusable library.0.2.0 — The ergonomic command-line framework for NURL. A single dependency-importable facade over the stdlib command-line surface (std/args parser, std/term isatty + AN0.1.0 — A training DataLoader in pure NURL — the batching layer the training stack was missing. Wrap an in-memory dataset (flat f64 features + optional labels) or strea0.1.5 — The embedding server, in pure NURL — load an XLM-RoBERTa-family text encoder (BGE-M3, multilingual-e5, …) from a Hugging Face model directory and serve embeddin0.3.0 — Bulletproof GGUF container toolkit — parse, verify, write and dequantise the ggml-ecosystem tensor format (llama.cpp, whisper.cpp, stable-diffusion.cpp) in pure0.11.0 — GPU compute for NURL — compiled kernels are cached on disk as device-specific CUBINs (keyed by source hash AND target architecture; loading a CUBIN needs no dri0.6.1 — The ergonomic GPU-compute toolkit for NURL — a diamond facade over the `gpu` package that kills the marshalling boilerplate every GPU-using package re-invents. 0.8.2 — Reverse-mode automatic differentiation over the tensor package — the training keystone. Build a computation eagerly with tape-recording ops (g_add, g_matmul, g_0.3.1 — The unified HTTP server interface for NURL — one dependency for anything that needs to SERVE HTTP. A single importable facade over the stdlib HTTP stack (socket0.1.1 — Fetch models from Hugging Face into one shared, verified, on-disk cache — the front door every model package was missing. Every model package used to start with0.1.1 — Isolation Forest anomaly detection — a CLI and a reusable library.0.5.0 — Pure-NURL image codecs + raster ops — decode, encode and transform images with no native library and no ImageMagick/ffmpeg shell-out. PNG decode covers the comp0.2.1 — LingBot-Map in pure NURL — the Geometric Context Transformer for streaming 3D reconstruction, ported from the reference PyTorch implementation. Feed it a folder0.1.2 — Render Markdown to HTML — a small CLI and reusable renderer library.0.3.5 — Trainable multi-layer perceptron in pure NURL — dense layers, ReLU, Adam, minibatches, early stopping. The training-side counterpart to the ecosystem's inferenc0.1.1 — Neural-network layers on the grad autograd tape — the reusable middle the training stack was missing. Every layer is a pure tape builder (GVar -> GVar over a GT0.1.1 — A tiny jq-lite JSON query tool: navigate, iterate, and pretty-print JSON from the command line.0.7.2 — Local MCP server for the NURL toolchain — exposes the installed compiler over Model Context Protocol so an LLM agent on the host can build, run, type-check, for0.12.11 — Run language models locally, in pure NURL — an ollama-shaped engine built package by package on the gguf container toolkit and the gpu compute package. `nurllam0.0.10.3.3 — Object detection from pure NURL, GPU-accelerated, using any tiny-YOLOv2-style ONNX model. Reads an image natively in any format the image package decodes (PNG, 0.7.3 — Run ONNX neural-network models from pure NURL, GPU-accelerated. The model's protobuf is decoded by a pure-NURL wire-format reader (no protoc, no external protob0.3.0 — Pure-NURL PostgreSQL client — v3 wire protocol, SCRAM-SHA-256 / MD5 auth, and TLS (pure-NURL, X25519 + P-256) with no libpq and no OpenSSL. Connects securely fr0.2.0 — Pure-NURL Redis client — RESP2 protocol over a libc TCP socket, with an optional pure-NURL TLS upgrade (rediss://, no OpenSSL on the target). A reusable library0.4.0 — The NURL package registry, served by NURL itself — a self-hostable registry speaking exactly the wire protocol nurlpkg already drives: GET /index/<name>.json + 0.3.1 — Read and write the safetensors tensor container in pure NURL — the format Hugging Face ships every modern model in — with a hostile-input-hardened reader and a 0.2.0 — A distributed compute cluster you join by installing it. Run `swarm worker <relay> <port>` on any machine and it announces itself over the relay group, every no0.21.0 — An MCP-controlled distributed compute engine with real GPU compute over real data. Every node runs the same command with composable role flags — --relay, --work0.1.0 — HTML template engine — Jinja-flavoured {{ vars }}, {% if/elif/else %}, {% for %} with loop.index/first/last, {% include %} partials and filters (raw/upper/lower0.4.3 — The reusable n-dimensional array for NURL — the shared ML layer over gpukit. It is the ndarray the rest of the ML stack builds on: onnx, grad, nn and anomaly si0.3.0 — The tokenizer that feeds a language model, in pure NURL — and it does not care which file the vocabulary arrived in. Two engines: SentencePiece-style bigram mer0.1.1 — Read PyTorch .pt / .pth checkpoints in pure NURL — the format every research model still ships in — without executing them. A torch.save file is a ZIP archive h0.1.0 — A vector index in pure NURL — the retrieval middle between embeddings and generation (the RAG piece). Build an index over n d-dimensional vectors and search the0.1.3 — Compile NURL to wasm32-wasi locally — no wasi-sdk, no build service. Drives the installed toolchain end to end: nurlc emits LLVM IR, the production IR rewriter 0.6.2 — A WebAssembly runtime written in pure NURL — load a wasm module and run it, with no external runtime. It decodes a wasm binary (all standard sections incl. the 1.0.5 — Speech recognition in pure NURL — whisper, running from the safetensors checkpoint Hugging Face ships, on the GPU or the CPU. `whisper transcribe openai/whisper0.6.6 — Promptable, open-vocabulary object detection AND instance segmentation from pure NURL on the GPU — a NURL port of YOLOE (THU-MIG, 'Real-Time Seeing Anything', I0.2.5 — Live YOLOE in the browser, served by pure NURL — open your webcam on a web page and watch open-vocabulary detection + instance segmentation stream back at you.