README.md

---
license: other
library_name: transformers
pipeline_tag: text-generation
tags:
  - soofi
  - mamba-2
  - moe
  - reasoning
  - thinking
  - sovereign-ai
  - preview
language:
  - en
  - de
  - es
  - fr
  - it
---

# Soofi-S-Isar-Preview Overview

> ⚠️ **Preview / internal checkpoint.** Weights and metadata may still change.

## Description

**Soofi-S-Isar-Preview** generates text responses with explicit chain-of-thought
reasoning before the final answer. It is one of two reasoning ("thinking")
variants of **SOOFI-S**, a sovereign, open-source language model developed by a
German research consortium, alongside
[Soofi-S-Rhine-Preview](https://huggingface.co/Soofi-Project/Soofi-S-Rhine-Preview).
SOOFI (Sovereign Open Source Foundation Models) is designed to provide a secure,
European open-source alternative to US and Chinese AI models for industrial use,
featuring strong reasoning and AI agent capabilities.

> TODO: describe how the Isar reasoning variant differs from Rhine
> (e.g. training data, reasoning style, target tasks).

This model is for research and development only (Preview).

## License/Terms of Use

Released under a custom license ("Other"). TODO: add the full license text /
link — the official card references a License section that is not yet filled in.

### Deployment Geography

Global (open release on the Hugging Face Hub). Development and training
infrastructure are located in Europe (see Computational Load).

### Use Case

Enterprise developers and researchers seeking a sovereign, European open-source
LLM for industrial use: tasks that benefit from explicit step-by-step reasoning
(math, logic, planning, complex analysis) and AI-agent / tool-use workflows.
English and German are the primary languages.

### Release Date

Hugging Face Hub — Preview at
<https://huggingface.co/Soofi-Project/Soofi-S-Isar-Preview>. TODO: final release
date (MM/DD/YYYY).

## Reference(s)

- Project: <https://soofi.info>
- Related models: see the *Related models* section below.
- TODO: link the technical report / paper once published.

## Model Architecture

**Architecture Type:** Transformer-based hybrid Mixture-of-Experts (MoE) with
Mamba-2 state-space (SSM) layers and attention layers. <br>
**Network Architecture:** Custom Hybrid Mamba-2/MoE (Nemotron-style), designed
from scratch — 23 Mamba-2/MoE layers + 6 attention layers; 128 routing experts
+ 1 shared expert per MoE layer; 6 experts activated per token. <br>

**This model was developed from scratch** (no base model). <br>
**Number of model parameters:** 3.0×10^10 total (30B), with ~3.5B active
parameters during inference.

## Computational Load

**Cumulative Compute:** TODO. <br>
**Estimated Energy and Emissions for Model Training:** TODO. Training
infrastructure is hosted entirely in Europe on T-Systems' Industrial AI Cloud
(Deutsche Telekom) to ensure data sovereignty.

## Input

**Input Type(s):** Text <br>
**Input Format(s):** String <br>
**Input Parameters:** One-Dimensional (1D) <br>
**Other Properties Related to Input:** Chat/ChatML-style messages via the
embedded chat template. No system prompt is required (none is injected by
default). Context length: see `config.json` (TODO: confirm maximum context).

## Output

**Output Type(s):** Text <br>
**Output Format(s):** String <br>
**Output Parameters:** One-Dimensional (1D) <br>
**Other Properties Related to Output:** As a reasoning model, Isar emits explicit
thinking traces (a `<think>` block) before the final answer; allow a generous
`max_new_tokens` budget. Supports the model's native tool-calling format.

## Software Integration

**Runtime Engine(s):** <br>
* Hugging Face `transformers` (`trust_remote_code=True`) <br>
* vLLM, llama.cpp/Ollama via the quantized variants (see *Related models*) <br>

**Supported Hardware Microarchitecture Compatibility:** <br>
* NVIDIA GPUs (Ampere and newer recommended) <br>

**Preferred/Supported Operating System(s):** <br>
* Linux <br>

The integration of foundation and fine-tuned models into AI systems requires
additional testing using use-case-specific data to ensure safe and effective
deployment.

## Model Version(s)

* **Soofi-S-Isar-Preview** — bf16 safetensors, unquantized (this repo). <br>
* Quantized derivatives: `…-GGUF` (llama.cpp/Ollama) and `…-FP8` (vLLM); see
  *Related models*.

## Installation & Usage

SOOFI-S ships with custom modeling code. You must load it using `trust_remote_code=True` with `transformers`.

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Soofi-Project/Soofi-S-Isar-Preview"
tok = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id, trust_remote_code=True, torch_dtype="auto", device_map="auto"
)

# No system prompt is required (none is injected by default).
messages = [{"role": "user", "content": "How many r's are in strawberry?"}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(**{"input_ids": inputs})   # sampling defaults come from generation_config.json

print(tok.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))
```

> As a reasoning model, Isar emits explicit thinking traces before the final
> answer; allow a generous `max_new_tokens` budget.

# Training, Testing, and Evaluation Datasets

### Dataset Overview

* **Total Size:** ~2.5×10^13 tokens (25 trillion). <br>
* **Languages:** English, German (primary); French, Italian, Spanish (limited).
  English acts as the pivot language. <br>
* **Knowledge Cutoff:** End of 2025. <br>
* **Training Start:** April 2026.

## Training Dataset

**Link:** TODO. <br>
**Data Modality:** Text. <br>
**Text Training Data Size:** More than 10 Trillion Tokens (~25T). <br>
**Data Collection Method by dataset:** Hybrid (freely available, high-quality
sources). TODO: refine. <br>
**Labeling Method by dataset:** TODO. <br>
**Properties:** Trained entirely from scratch on freely available, high-quality
tokens.

## Testing Dataset

**Link:** TODO. <br>
**Properties:** TODO.

## Evaluation Dataset

**Link:** TODO. <br>
**Benchmark Score:** TODO — add key reasoning benchmarks once available. <br>
**Properties:** TODO.

## Inference

**Acceleration Engine:** `transformers`; vLLM / llama.cpp via quantized
variants. <br>
**Specific Test Hardware:** TODO.

## Ethical Considerations

The SOOFI consortium believes Trustworthy AI is a shared responsibility and has
established policies and practices to enable development for a wide array of AI
applications. When downloaded or used, developers should work with their
internal model team to ensure this model meets requirements for the relevant
industry and use case and addresses unforeseen product misuse.

For more detailed information, see the Model Card++ subcards below. Please report
model quality, risk, security vulnerabilities, or concerns to
<a href="mailto:contact@soofi.info">contact@soofi.info</a>.

### Bias Subcard

| Field | Response |
|:---|:---|
| Participation considerations from adversely impacted groups in model design and testing | TODO |
| Measures taken to mitigate against unwanted bias | TODO |
| Bias Metric (if measured) | TODO |

### Explainability Subcard

| Field | Response |
|:---|:---|
| Intended Task/Domain | Reasoning-heavy tasks (math, logic, planning, analysis), AI-agent/tool use |
| Model Type | Hybrid Mixture-of-Experts (MoE) autoregressive reasoning ("thinking") language model |
| Intended Users | Enterprise developers and researchers |
| Output | Text (String), with an explicit `<think>` reasoning trace before the answer |
| Describe how the model works | Generates text autoregressively; a router activates 6 of 128 experts per token across hybrid Mamba-2/MoE and attention layers; emits chain-of-thought before the final answer |
| Technical Limitations | Preview checkpoint; non-primary languages (FR/IT/ES) are limited; reasoning traces can be verbose; may produce inaccurate or outdated content (knowledge cutoff end of 2025) |
| Verified to have met prescribed quality standards | TODO |
| Performance Metrics | TODO (see Evaluation Dataset) |
| Potential Known Risks and Mitigation | May generate incorrect, biased, or unsafe content, including flawed reasoning; apply use-case-specific testing and guardrails before deployment |
| Terms of Use/Licensing | Other (see License/Terms of Use) |

### Privacy Subcard

| Field | Response |
|:---|:---|
| Generatable or reverse engineerable personal data? | TODO |
| Personal data used to create this model? | TODO |
| Was consent obtained for any personal data used? | TODO |
| How often is dataset reviewed? | TODO |
| Was data from user interactions with the AI model used to train the model? | No |
| Is there provenance for all datasets used in training? | TODO |
| Applicable Privacy Policy | TODO |

### Safety & Security Subcard

| Field | Response |
|:---|:---|
| Model Application Field(s) | Industrial use; customer service; reasoning, planning, and agent applications |
| Describe the life critical impact (if present) | None intended. Not for use in life-critical or safety-critical decision-making without independent validation |
| Use Case Restrictions | Abide by the applicable license agreement (see License/Terms of Use) |
| Model and dataset restrictions | TODO |

## Related models

- Instruct variant: [Soofi-Project/Soofi-S-Instruct-Preview](https://huggingface.co/Soofi-Project/Soofi-S-Instruct-Preview)
- Other reasoning variant: [Soofi-Project/Soofi-S-Rhine-Preview](https://huggingface.co/Soofi-Project/Soofi-S-Rhine-Preview)
- GGUF quantizations (llama.cpp/Ollama): [Soofi-Project/Soofi-S-Isar-Preview-GGUF](https://huggingface.co/Soofi-Project/Soofi-S-Isar-Preview-GGUF)
- FP8 quantization (vLLM): [Soofi-Project/Soofi-S-Isar-Preview-FP8](https://huggingface.co/Soofi-Project/Soofi-S-Isar-Preview-FP8)

## Citation

```bibtex
@misc{soofi_s_isar_preview,
  title  = {Soofi-S-Isar-Preview},
  author = {SOOFI Consortium},
  year   = {2026},
  url    = {https://huggingface.co/Soofi-Project/Soofi-S-Isar-Preview}
}
```