Sabi-Noema

Sabi-Noema

A 27B agentic text model built on Qwen3.8-27B with open-source RL environments, reporting higher terminal and coding-agent scores, for developers building coding and tool-using agents.

Made by sabicap 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: Sabi-Noema is an agentic text model from sabicap, built on top of Qwen3.8-27B. The authors say they trained it only with open-source reinforcement learning environments, aiming at stronger small-model skill with less wasted output. It has about 27B parameters in a dense hybrid design, a 262K-token context and BF16 weights of roughly 55 GB. The card dates the release to October 2026. A separate variant, Sabi-Noema-MLA, uses multi-head latent attention.

Who it suits: It suits developers who build coding and tool-using agents and want a 27B open model that the authors say finishes terminal and software tasks with fewer output tokens.

Title: Who uses this model
Industry | Who uses it | How they use it | Example
Software | Developers | Run a coding agent that works in a terminal | Let it handle a multi-step repository task
AI tooling | Agent builders | Drive tool calls over many turns | Call it from a support or retail workflow
AI research | Researchers | Compare an agent fine-tune with its base | Check its scores against Qwen3.8-27B

If you need a small model for a basic laptop, you can skip it, since the BF16 weights alone are about 54.7 GB.

Where to use it:

• Python scripts that use the transformers library
• The vLLM server, which the card lists as supported in a stock build
• llama.cpp, after a GGUF conversion
• Windows, macOS or Linux with Python, for downloading the files

Title: Jobs it does
Use case | Who | What they get back
Terminal and coding tasks | Developers | An agent that works through multi-step software work
Tool calling | Agent builders | Calls across several turns of a conversation
Lower hallucination | Teams that need dependable answers | A lower rate on the card's own test than the base model
Fewer output tokens | Cost-conscious teams | Shorter outputs on tasks both models solve, per the authors

How to set it up: You need Python and the huggingface_hub package, plus a GPU setup that can hold a 27B model. The weights are about 54.7 GB in BF16, or about 13.7 GB at 4-bit by Hugging Face's estimate.

• Python with pip
• About 54.7 GB of free disk space for the BF16 files
• A GPU or server setup that can hold a 27B model
• The transformers library, vLLM or llama.cpp

1. Install huggingface_hub.
2. Download the model files with the hf command.
3. Load them with your chosen library and the generation settings the card points to.

pip install -U huggingface_hub
hf download sabicap/sabi-noema

How to use it: The card says to use the same generation settings as Qwen3.8-27B, which are the defaults saved in the model's generation configuration file. The chat template also supports turning thinking off through the enable_thinking option, but the authors did not test that mode. The visible part of the card shows no further run steps, so check the full card for loading code.

Title: Main commands
Tool | What it does
pip install -U huggingface_hub | Installs the helper that downloads model files
hf download sabicap/sabi-noema | Saves the model files on your computer

Example: You download the files, serve them with vLLM and give an agent a task in a terminal. By the authors' account, the agent finishes it with fewer output tokens than the base model would use on tasks both solve.

For newcomers and experts: An agent model is meant to take actions, not just chat. A first small step is to compare your memory against the 13.7 GB 4-bit estimate and the 54.7 GB BF16 size before you download.

The card describes training that starts from Qwen3.8-27B and uses reinforcement learning environments that are all open source. Its tables use the same test setup for every model, with Terminal-Bench 2.1 run on H100 setups and the other tests described in the card. These are the authors' own numbers, so treat them as claims to check.

Qwen3.8-27B base model
Open-source RL environments
Sabi-Noema agent model
Terminal and tool tasks

Good to know: The authors report 76.0% on Terminal-Bench 2.1 against 73.0% for the base, 47.5% on DeepSWE against 38.5%, and a hallucination rate of 15.1% against 30.4%. They also say it uses about 2.3 times fewer output tokens than the base on Terminal-Bench tasks both solve. Other people have published 9 quantized and 3 fine-tuned versions.

Title: Terminal-Bench 2.1 score, in percent
Sabi-Noema: 76.0
Qwen3.8-27B: 73.0
GPT-6 Luna, high: 68.5
ThinkingCap: 67.8
Swift 1.5: 62.9
Muse Glimmer 30B: 48.3

• License: Apache 2.0.
• Cost to run: none named, apart from the hardware.
• Data it reaches: the files come from Hugging Face, and the model then runs on your own computer.
• Not for: small machines, since the files are large.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 27.4B
• Weights: about 54.7 GB in BF16; about 13.7 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,810
• Likes: 55
• License: apache-2.0
• Based on: Qwen/Qwen3.8-27B
• Quantized versions: 10
• Fine-tunes: 3
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, agentic, reinforcement-learning, tool-use, hybrid-attention, qwen3.8, conversational, text-generation

Links

• Model: https://huggingface.co/sabicap/sabi-noema
• Files: https://huggingface.co/sabicap/sabi-noema/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download sabicap/sabi-noema

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