Directory
Published directory entries matching your search.
A directory of abstracts and links to academic papers focused on local language models, current as of the end of March 2026. The listing compiles research published between 2017 and 2023, including foundational works such as Attention Is All You Need, BERT, T5, and PaLM, alongside studies on efficient transformer inference, multimodal models, and instruction tuning.
Magic.dev is an artificial intelligence company focused on developing safe artificial general intelligence (AGI) to accelerate progress on high-priority global problems. The company builds frontier code models intended to automate software engineering and AI research workflows. Its technical approach combines large-scale pre-training, domain-specific reinforcement learning, extended context windows, and increased inference-time compute. Magic.dev employs a small team of engineers and researchers working on fundamental AGI-related research problems.
Nexusflow/NexusRaven-V2-13B is a 13-billion parameter language model available on Hugging Face. It is designed for zero-shot function calling and text generation, aiming to advance open-source artificial intelligence. The model can be used with libraries such as Transformers and provides instructions for deployment via inference providers, notebooks, and local applications.
NVIDIA AI is a suite of enterprise artificial intelligence solutions offered by NVIDIA. The portfolio includes accelerated infrastructure, enterprise-grade software, and AI models. Solution areas span agentic AI, inference, conversational AI, data science, and cybersecurity. The program aims to help organizations transform into AI-enabled enterprises.
NVIDIA NIM is a catalog of models offered as inference microservices for deployment on the user's GPU infrastructure. The listings include downloadable models and free or partner endpoints, organized by use cases such as image generation, speech-to-text and drug discovery.
OmniRoute by Cheaper Inference is an open-source AI router that routes requests across providers through one OpenAI-compatible endpoint, with automatic fallback. It accepts OpenAI, Claude, Gemini and Responses API requests at /v1 and ships as a self-hosted gateway with a local dashboard. The catalog covers chat, media, search, local runtimes and cloud agents.
NeMo Parakeet ASR Models attain strong speech recognition accuracy while being efficient for inference. Available in CTC and RNN-Transducer variants.
PocketPal AI is an open-source app for running language models directly on a device, available on iOS and Android. It supports GGUF models from Hugging Face, on-device processing, device benchmarking, conversation history and adjustable inference settings.
Python Library for Probabalistic Programming (Bayesian Inference and Machine Learning).
Radiata is a stable diffusion web UI built on the diffusers library that uses TensorRT high-speed inference technology. Its official documentation provides resources for introduction, installation, usage, developer guidance, and troubleshooting for Windows and Linux operating systems. It is categorized under AI & Machine Learning with a focus on AI image generation.
Server Survival is a free, open-source browser game that teaches cloud architecture through 25 campaign levels in 11 languages, without installation. It covers topics such as load balancing, caching, autoscaling, circuit breakers, multi-region failover and GPU inference economics.
SmolVLA is a compact 450M open-source Vision-Language-Action model for robotics that runs on consumer hardware. It is pretrained on compatibly licensed, open-source community-shared datasets under the lerobot tag and supports asynchronous inference.
Stable Beluga 1 is a text generation model developed by Stability AI, released under a CC-BY-NC-4.0 license. It can be used with libraries such as Transformers, via inference providers, in notebooks, or in local applications.
Stable Beluga 2 is a text generation model developed by Stability AI. It can be used with libraries such as Transformers, via inference providers, in notebooks, or in local applications.
Stan is software for Bayesian data analysis that combines statistical modeling capabilities with user-friendly interfaces and open-source development. Its probabilistic programming language supports applications from simple linear regression to multi-level models and time-series analysis. Interfaces exist for Python, Julia, R, and the Unix shell, usable on laptops, clusters, or in the cloud. Official documentation covers the language, inference algorithms, and built-in functions.
tRPC lets developers build and consume fully typesafe APIs without schemas or code generation. It uses TypeScript type inference and has zero runtime dependencies. Adapters support frameworks including React, Next.js, Express, and Fastify.
Velda is a serverless GPU cloud service for running training and inference jobs without container or cluster management. A vrun prefix runs commands from the local development environment, and compute is provisioned only while jobs run. Pricing is pay-as-you-go, with listed GPU and CPU pools billed per hour and optional self-hosted or enterprise deployment.
vLLM is an open-source library for efficient serving of large language models. It provides an API server, supports offline and online inference, deployment options, and features such as automatic prefix caching, batch invariance, context extension, and speculative decoding. It can be used with GPUs, CPUs, and TPUs.
WebNN is a web standard that lets web apps and frameworks accelerate deep neural networks using on-device hardware such as GPUs, CPUs, or purpose-built AI accelerators. The introduction states that in-browser inference supports local media sources, keeps user data on-device, and needs no network after initial asset caching.
A Hackaday article by Matthew Carlson explaining why LLaMa matters. It describes LLaMa as a transformer language model from Facebook/Meta research spanning 7 billion to 65 billion parameters, and notes that llama.cpp runs inference with 4-bit quantization.
Bittensor is an open, decentralized network where independent subnets produce digital commodities such as compute, inference, storage and prediction, and the chain pays contributors in TAO. Miners produce the commodity, validators score miners, subnet creators define the incentive mechanism, and stakers back validators. The Bittensor Python SDK and btcli cover user-facing chain operations.
