What is Multi-Stream LLMs?
ELLIS Tübingen research code for training LLMs that read, think and write in parallel streams.
Who is Multi-Stream LLMs best for?
Multi-Stream LLMs is best for language model researchers who want to reproduce experiments on models that think in parallel streams.
Researchers testing new model formats
It gives each role, such as user input, answer and thinking, its own stream, so one forward pass reads and writes several streams at once.
Efficiency experiments
One folder trains Qwen3 1.7B and 4B models to solve a problem while still reading it, with math, logic and QA benchmarks.
Prompt-injection research
Another folder trains Qwen2.5 7B and Qwen3 4B models with separate streams and tests them on TensorTrust, Gandalf and other injection benchmarks.
Monitoring model reasoning
A third folder trains 8B and 27B models with eight thinking streams to study how well a monitor can read their reasoning.
How much does Multi-Stream LLMs cost?
Multi-Stream LLMs is free research code on GitHub with no paid tier, but the repository has no license file, so its reuse terms are not stated.
- Training and evaluation code from the paper
- No license file in the repository
- No trained model weights published
The code comes from the SEAL group at the ELLIS Institute Tübingen and the Max Planck Institute for Intelligent Systems as the release for their Multi-Stream LLMs paper. It builds on Qwen base models, and each of its three folders sets up and runs on its own. Full pricing on github.com. We checked these prices on .
How do readers rate Multi-Stream LLMs?
No one has rated Multi-Stream LLMs yet. Be the first.
Sign in to rate. One rating per account.
Questions about Multi-Stream LLMs
What is Multi-Stream LLMs?
ELLIS Tübingen research code for training LLMs that read, think and write in parallel streams.
Who is Multi-Stream LLMs for?
Multi-Stream LLMs is best for language model researchers who want to reproduce experiments on models that think in parallel streams.
Is Multi-Stream LLMs free?
Yes. Multi-Stream LLMs has a free plan: training and evaluation code from the paper.
Where can I try Multi-Stream LLMs?
On its official website, github.com.







