WERM

WARP (Weight-Aware Runtime and Paging), formerly WASTE, is an embeddable inference engine written in dependency-free C by SQLiteAI. It runs large mixture-of-experts models beyond available RAM by keeping the shared model trunk in memory, streaming selected experts from NVMe or other fast storage, and using remaining RAM as a bounded expert cache. Its container layout gives each expert an aligned read; reads overlap computation, while a lookahead router prefetches experts for the next layer without changing the decisions made by the actual router. Experts use 3-bit residual vector quantization, while more sensitive shared weights use 4- or 8-bit storage. The project is designed for local inference of frontier-scale models such as Kimi K3 and GLM-5.3-Flash on consumer hardware.

WebLLM is an in-browser large language model inference engine developed by the MLC community. It runs model inference directly in web browsers using WebGPU hardware acceleration, without server-side processing, and can be used as an npm, Yarn, or CDN package for building web applications. Its MLCEngine exposes an OpenAI-compatible chat-completions interface with streaming, JSON-mode structured generation, seeding, and preliminary function-calling support. It supports model families including Llama, Phi, Gemma, Mistral, and Qwen, and can load custom MLC-format models consisting of model artifacts and WebAssembly model libraries. WebLLM provides Web Worker and Service Worker integrations, Chrome-extension examples, browser caching through Cache API, IndexedDB, or OPFS, and optional Subresource Integrity verification for downloaded model artifacts.

NeoHorse is TokenRhythm's family of open-weight causal language models for text-based agent workflows, including tool use, coding, and instruction following. NeoHorse-1 uses routing-guided agentic post-training: a harness assigns tasks to a heterogeneous model pool, records tool interactions and outcomes, estimates capability demand, and feeds capability-level feedback into later training mixtures. Updated models can return to the harness in an evaluation-selection-update loop intended as a prototype for recursive self-improvement. NeoHorse-1 is released in approximately 4B and 9B parameter variants, post-trained from Qwen3.5 models, with text input and text output interfaces and deployment examples for vLLM and SGLang. The related NeoHorse-Jev-4B model uses prefill-only inference to return structured Choice, Noul, and Score decisions and probabilities for routing requests, selecting tools, or controlling workflows. The models are released under the Apache License 2.0.

WormGPT is a named AI chatbot reported to be designed and marketed for cybercriminal use, offering guidance and code related to malware, phishing, and intrusion techniques. It emerged in reporting from late 2023 and was discussed by cybersecurity firms and in underground forums as an illicitly purposed generative AI service.

NanoJev is an open-source 0.6B parallel-decision model and training pipeline based on Qwen3-0.6B. It accepts states, questions, and candidate sets in a single forward pass and returns complete probability distributions without output-token decoding. Shared decision heads support dynamic-choice distributions over supplied candidates, Boolean proposition probabilities, and ordered score levels with expected values. The repository includes data generation, distribution-loss training, evaluation, persistent serving, and browser replay tools. Its game demonstrations combine model judgments with controller code: maze controllers ask local safety questions and track collisions and explored paths, while Snake controllers filter immediate collisions, plan toward visible food, and use the model to break ties between remaining actions. Checkpoints and datasets are distributed through Hugging Face, and the service exposes repeated batched evaluations through a local HTTP endpoint.