Tev-320B

Tev-320B

A 320B mixture-of-experts agent model that works through long terminal tasks with a single bash tool. Trained with reinforcement learning on 1,170 environments, for researchers under a non-commercial license.

Made by gordonhu 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: Tev-320B is an agent model for long terminal tasks. It starts from GLM-5.3-Flash, a mixture-of-experts model with 320B total and 18B active parameters. It was trained with reinforcement learning on 1,170 verified environments that grade every step, not only the final result.

Who it suits: It suits researchers who study terminal agents and long-horizon tasks, and teams that can host a huge model. The license is non-commercial.

Title: Who runs it
Industry | Who uses it | How they use it | Example
Research | Agent researchers | Study reinforcement learning with progress rewards | Compare Tev-320B with its GLM base
Software | Teams testing coding agents | Run an agent that works only through bash | Repair a broken build in a container
Evaluation | Benchmark teams | Run terminal tasks and score pass rates | Score it on Terminal-Bench subsets

If you need a small model or a commercial product, you can skip it.

Where to use it:

• The transformers library, on Windows, macOS or Linux with Python
• Terminal agent frameworks that give the model one bash tool in a container
• Multi-GPU servers, since the weights are about 643 GB in BF16

Title: Jobs it does
Use case | Who | What they get back
Long terminal tasks | Agent builders | Step-by-step bash work toward a goal
Environment repair | Researchers | A workspace fixed turn by turn
Benchmark runs | Evaluators | Pass rates on Terminal-Bench subsets

How to set it up:

• Python on Windows, macOS or Linux
• Large storage and multi-GPU memory, since the files are 49 safetensors shards of about 643 GB
• An agent setup that hands the model one bash tool inside a container made for each task

1. Install the Hugging Face Hub tools.
2. Download the model files.
3. Load the model in your agent framework and give it the bash tool.

pip install -U huggingface_hub
hf download gordonhu/Tev-320B

How to use it: Run the model as an agent. It receives a terminal task in a container and answers with bash commands, turn by turn. The card describes this one-tool setting but shows no run commands.

hf download gordonhu/Tev-320B

Example: You place the agent in a container that holds a broken data pipeline and ask it to fix the pipeline. It works through bash commands one turn at a time and its workspace gets closer to the expected state. The card does not show a sample session.

For newcomers and experts: If you are new, this is an AI that works at a command line, running commands and reading the results until a task is done. The first small thing to try is reading the card's description of how Tev trains its agents.

If you know the field, training uses a latent world, a hidden correct specification that grades both the final answer and every step. The reward adds 0.25 times a progress score to a pass or fail check. The trade-off is evidence: results come from the author's own subsets of Terminal-Bench, with 22 tasks in the 3.0 set and 18 in the 4.0 set.

Build a task from a hidden correct world
Corrupt the inputs with repairable defects
The agent works in a container with bash
Score each turn by progress, then the result

Good to know:

Title: Terminal-Bench 4.0 CM subset, Pass@1
Claude Opus 5: 0.52 Pass@1
Claude Fable 5: 0.46 Pass@1
GPT-5.6 (sol): 0.39 Pass@1
Tev-320B: 0.37 Pass@1
Grok 4.6: 0.31 Pass@1
GLM-5.3-Flash: 0.29 Pass@1
Claude Opus 4.8: 0.27 Pass@1
Claude Sonnet 5: 0.19 Pass@1

• Gain over its base: Pass@1 is 0.27 against 0.20 on the 3.0 subset, and 0.37 against 0.29 on the 4.0 subset.
• Domains: Tev-Gym covers seven domains, from software engineering to quantum computing.
• Weights: the files carry the base model's vision encoder untouched. Training and tests used text only, and no MTP weights ship with it.
• License: cc-by-nc-4.0, so it is non-commercial.
• Cost to run: no API key or paid service is named, but the model needs a huge amount of GPU memory.
• Data it reaches: the agent runs bash in a container, so what it can reach depends on how you set that container up.
• Not for: commercial use, which the license bars.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 321B
• Weights: about 643 GB in BF16; about 161 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 495
• Likes: 17
• License: cc-by-nc-4.0
• Languages: en
• Based on: zai-org/GLM-5.3-Flash
• Quantized versions: 2
• Tags: transformers, safetensors, glm5_next, image-text-to-text, agent, long-horizon, terminal, reinforcement-learning, text-generation, conversational

Links

• Model: https://huggingface.co/gordonhu/Tev-320B
• Files: https://huggingface.co/gordonhu/Tev-320B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download gordonhu/Tev-320B

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