Terminal-RSI-27B

Terminal-RSI-27B

Works through multi-step terminal tasks such as inspecting a repository, editing code and repairing dependencies. A 27.8B-parameter agent model built on Qwen3.8-27B for researchers and agent builders.

Made by Teriminal-RSI on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Terminal-RSI-27B is a terminal-agent model from the Terminal-RSI project, built on Qwen3.8-27B. It is meant for tasks that need a running environment and a check of the result, such as repository inspection, code edits, dependency repair and rebuilding scientific workflows. Its training alternates two steps the card calls Recursive Self-Rewrite and Recursive Policy-Refine.

Who it suits: Researchers and developers who build or test agents that work in a terminal and check their own results.

Industry | Who uses it | How they use it | Example
Software development | Agent builders | Run multi-step coding tasks in a terminal | Fixing a broken dependency in a repository
Research | AI researchers | Study terminal-agent training and evaluation | Comparing runs on a terminal benchmark
Science | Workflow engineers | Rebuild a scientific workflow step by step | Reconstructing an analysis pipeline
Engineering tools | Tool developers | Inspect and edit code through commands | Reading a repository and changing a file

If you want a plain chat model, you can skip it, because the card says a standalone chat example does not reproduce its agent results.

Where to use it:

• An agent runner that gives the model a terminal, tools and a checker
• Python with the transformers library, after you download the files
• Your own Windows, macOS or Linux machine with enough memory

Use case | Who | What they get back
Repository inspection | Developers | Commands run and findings about the code
Dependency repair | Developers | Edits and reruns until the environment works
Workflow reconstruction | Scientists | A rebuilt multi-step workflow checked by a verifier

How to set it up:

• Python and pip
• About 55.6 GB of disk space in BF16, or about 13.9 GB at 4-bit (an estimate)
• An agent runner and a safe environment where it can run commands

1. Install the Hugging Face hub tools.
2. Download the model files.
3. Load the model by following the model card, then connect it to your runner.

pip install -U huggingface_hub
hf download Teriminal-RSI/Terminal-RSI-27B

How to use it: Connect the model to a runner that lets it act in a terminal and checks the state afterwards. The paper's tests used a temperature of 0.60, a top-p of 0.95, JSON actions and thinking mode turned off. Results depend on the runner, tools, prompts and time budget.

Example: You give the agent a repository with a broken dependency and a task to get its tests passing. It inspects files, edits the dependency setup, reruns commands and stops when the verifier confirms the state.

For newcomers and experts: If you are new to agents, treat this as research material rather than a ready tool. Start by reading how the card describes its tasks before you connect it to a terminal.

It has 27.8B parameters, 64 text layers, a hidden size of 5,120 and a configured context of 262,144 tokens. Training follows a base, RSR, RPR, RSR, RPR sequence, with PPO in the refine steps. The card marks its results as the paper's own, from the retained RPR-2 checkpoint.

Agent gets a task in a terminal
It runs commands and edits files
A verifier checks the resulting state
It retries until the task passes

Good to know: Benchmarks: the paper reports Terminal-Bench scores, averaged over three runs, and the card omits outside model comparisons

Title: Success rate in percent, base model against Terminal-RSI-27B, paper-reported
Terminal-Bench 2.1 base: 51.70%
Terminal-Bench 2.1 Terminal-RSI: 86.52%
Terminal-Bench Hard base: 31.00%
Terminal-Bench Hard Terminal-RSI: 79.00%
Four-suite mean base: 20.68%
Four-suite mean Terminal-RSI: 50.58%

• License: apache-2.0
• Cost to run: no API key or paid service is named, only hardware for a 27.8B model
• Data it reaches: an agent run reads and changes files in whatever environment you give it, so use a sandbox
• Not for: casual chat, since the card says a standalone chat example does not reproduce its agent evaluations

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 27.8B
• Weights: about 55.6 GB in BF16; about 13.9 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 2,536
• Likes: 33
• License: apache-2.0
• Languages: en, zh
• Quantized versions: 5
• Fine-tunes: 1
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, terminal-rsi, terminal-agent, reinforcement-learning, recursive-self-improvement, code, long-context, text-generation, conversational

Links

• Model: https://huggingface.co/Teriminal-RSI/Terminal-RSI-27B
• Files: https://huggingface.co/Teriminal-RSI/Terminal-RSI-27B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Teriminal-RSI/Terminal-RSI-27B

Then follow the usage in its model card. Nothing in this kit runs by itself.