Publication window: 2024–2025.
Qwen3 combines fast responses and extended reasoning in one model family, with a configurable thinking budget and multilingual training.
Connects your interest in language models with how information is represented and retrieved.
The study explores reinforcement learning for language-model reasoning and transfers reasoning capabilities to smaller models through distillation.
Connects your interest in language models with how information is represented and retrieved.
A small curated training set and budget forcing let a language model spend more computation checking its answers on reasoning tasks.
Connects your interest in language models with how information is represented and retrieved.
This report describes a mixture-of-experts language model and the architecture and training choices used to improve computational efficiency.
Connects your interest in language models with how information is represented and retrieved.
OLMo 2 releases model weights, training data, code, and checkpoints, alongside techniques for more stable and efficient language-model training.
Connects your interest in language models with how information is represented and retrieved.
The Llama 3 report documents a family of language models, covering pretraining, post-training, and evaluation across a range of language tasks.
Connects your interest in language models with how information is represented and retrieved.
Discover papers that fit your interests.
Product framework
Explore the research loop and how each step connects.
Discover academic papers matched to your research interests, with clear reasons for each recommendation.
Explore the feedTurn papers and notes into a personal wiki. Follow the links between ideas, with their sources close by.
Explore the wikiThink alongside Sparky. Bring the papers you have read into new questions worth exploring.
Explore Idea SparkCreate an account and open your research workspace. Your vault lives online, ready when you return.
Start with an included AI allowance. You can also connect your own API key.
Already have a workspace? Sign in
Run SciSpark locally with a Markdown vault you control.
Requires Node.js 20.9+ and npm.
git clone https://github.com/SciSpark-ai/scispark_wiki.git
cd scispark_wiki
npm install
npm run devThen open 127.0.0.1:3000
For AI, connect an API provider or a signed-in Codex or Claude Code CLI.
Setup guideSciSpark is an AI workspace for academic research across disciplines. Discover scholarly papers, read with context, and build connected knowledge. It brings your paper feed, personal wiki, knowledge graph, and research ideas into one place, with Sparky as your research companion.
The hosted beta runs in your browser and keeps your vault online. Local installation stores your Markdown vault on your computer. Both approaches may send requests to scholarly services and your chosen AI provider. Automatic sync between hosted and local vaults is not currently part of this offer.
The hosted beta includes a starter AI allowance, and you can choose to connect your own API key. Local installation needs an API provider or a signed-in Codex or Claude Code CLI for AI features. Provider charges depend on the option you use.
SciSpark uses your research interests and profile to find academic papers from scholarly sources such as arXiv, OpenAlex, Semantic Scholar, and PubMed. In the preview above, choose one of three topics to see how discovery connects to reading, notes, and ideas. The preview is not a live recommendation feed.
A native mobile app is planned, so your research can be closer at hand. For now, you can open the hosted beta in your browser.
Discover with purpose. Read with context. Keep what you learn.