Powerful system-level package manager for Linux, macOS and Windows written in Rust β building on top of the Conda ecosystem.
Frontier-class open models on a free Kaggle TPU v5e-8: GLM-5.3-Flash 320B MoE (~165 tok/s, our own JAX engine) and Qwen3.8-27B bf16 (~130 tok/s), 262k context, prefix caching. Works with Claude Code, Codex, opencode and pi.
Installs the latest UMU/GE-Proton and Non Steam Launchers under 1 Proton prefix folder and adds them to your steam library. Installs... Battle.net, Epic Games, Ubisoft, GOG, EA App, Amazon Games, itch.io , Legacy Games, The Humble Games Collection, IndieGala, Rockstar Games Launcher,PS Plus, Glyph, VKPlay & HoYoPlay, Waydroid and Games on SD Card.
Collection of dozens of minimal bare-metal x86 operating system examples for learning low-level system programming, tested in QEMU and on real hardware.
Run Windows software and games on Linux
DeepSeek Harness η ΄η²ε’εΌΊη (enhanced fork): TUN fake-ip dynamic pass-through (patch #38), persona prefix variants, Node 24 compat, idempotent anchors
A clean, sleek prefix and runtime manager for Rosemary
Hermes Agent context engine: the main model writes the compaction handoff on its cached prompt prefix
vMLX - Use MLX models easily - JANGQ (GGUF for MLX) - Not dependant on mlx_vlm
GLM-5.3-Flash EXL3 (320B MoE) on 2x NVIDIA DGX Spark β production serving kit, 1M context, 97%+ multi-session prefix caching, DFlash2 spec decode
ZipTok3D: High-Fidelity 3D Tokenization with Compact Token Prefixes
π¦ Share target browsers between different front-end tools, like Autoprefixer, Stylelint and babel-preset-env
FilterIterator implementation that filters files based on a list of suffixes, prefixes, and other exclusion criteria.
A SOAP client and server for node.js.
Autoprefixer for Ruby and Ruby on Rails
OpenCL integration for Python, plus shiny features
π¨ Prefixes all PHP namespaces in a file/directory to isolate the code bundled in PHARs.
Data structure and relevant algorithms for extremely fast prefix/fuzzy string searching.
Sublime plugin to prefix your CSS
A code generator for array-based code on CPUs and GPUs
PruningRadixTrie - 1000x faster Radix trie for prefix search & auto-complete
An open training and inference stack for Jev-style decision models. Train models to score dynamic candidate branches from a shared prefix, with support for high-cardinality choice, calibration, and fast batched inference.
Coproton 2026: Run Steam Games and Trainers Together in One Proton Prefix
A minimal coding-agent harness built to measure what context management actually costs: constraint survival after compaction (8%β83%), default-FAIL verification against false completions, and KV-cache prefix design (21%β90%). Stdlib-only, real deepseek-chat runs
Streaming speech recognition toolkit: log-mel features, CTC greedy and prefix-beam decoding, transducer-style frame-synchronous emission, forced alignment, Chinese/English CER-WER metrics, and a chunked streaming pipeline with latency accounting. NumPy core with a CPU-only torch extra; trainable demos on synthetic audio, fully offline.
Prefix Sliding for efficient test-time scaling
IP Subnet Calculator β Enter any IPv4 address with a CIDR prefix or subnet mask to instantly get the network address, broadcast address, usable host range, wildcard mask, binary/hex breakdown, and classification. Includes a VLSM planner, subnet list generator, and reverse calculators for required hosts or subnets. Free on ToolAdda.online.
Checkers to test your internet provider for censorship.
ASN / RPKI validity / BGP stats / IPv4v6 / Prefix / URL / ASPath / Organization / IP reputation / IP geolocation / IP fingerprinting / Network recon / lookup API server / Web traceroute server
Bulk domain availability checking via RDAP, DNS, and WHOIS, with prefix/suffix permutations, regex patterns, MCP, and multiple output formats
Open-source TypeScript terminal coding agent for DeepSeek-V4 β builds on DeepSeek's strong price-performance and ultra-cheap cache pricing, engineering byte-stable prefixes and cache-reusing forks so cross-session memory and a continuous self-correction layer add almost no token cost; 1M context, Skills/MCP/Hooks, Claude Code config compatible.
A Mini-vLLM optimization project for Qwen3-32B TP=2 serving, Prefix Cache, and Triton PagedAttention.
Faster inference backend for Open-Jev-27B: fused CUDA kernels, prefix tree, CUDA Graphs (B300, bf16)
Serving Qwen3.8-27B-FP8 on a single DGX Spark (GB10): 7.88 to 58.5 tok/s single-stream from decode strategy alone, weights untouched. Speculative decoding and prefix caching benchmarked, plus DFlash 2 β the only Qwen3.8-27B build that can serve it under vLLM.
Parse CSS and add vendor prefixes to rules by Can I Use