README.md

---
license: mit
language:
- en
- hi
- zh
- it
library_name: transformers
tags:
- SPEED
- SKT
- SKT AI LABS
---

### SPEED 1

<div align="center" style="font-family: 'Segoe UI', Tahoma, Geneva, Verdana, sans-serif; max-width: 1000px; margin: 0 auto; padding: 20px;">

  <div style="margin-bottom: 15px;">
    <a href="https://sktailabs.in">
      <img src="https://www.sktailabs.in/skt-icon-96.png" alt="SKT AI LABS Logo" style="width: 96px; height: 96px; object-fit: contain;" />
    </a>
  </div>
  <a href="https://sktailabs.in" style="text-decoration: none;">
    <h1 style="color:#1b4fe8; font-size:48px; font-weight:900; text-shadow: 2px 2px 6px rgba(0,0,0,0.2); cursor: pointer; margin: 0 0 15px 0;">SKT AI LABS</h1>
  </a>
</div>

![SPEED](https://cdn-uploads.huggingface.co/production/uploads/6970b3cd48a639796bf613ac/AMgzYhdGjuBJLGBZqg6G9.png)

**SPEED-1** is a Mixture-of-Experts (MoE) language model with ~19B parameters, built by SKT AI LABS Private Limited, an Indian AI company based in Sidhi, Madhya Pradesh. It is designed for complex tasks across a wide range, trained for complex reasoning, with chain-of-thought data spanning agentic tasks, mathematics, and coding.

---
## Evaluation Results

Early results from ongoing development (October 2026, evaluated on AMD MI300X). Detailed eval configuration will be published with the full release.

<div align="center">
<table>
<thead>
<tr>
<th align="center" style="text-align: center">Benchmark</th>
<th align="center" style="text-align: center">SPEED-1</th>
</tr>
</thead>
<tbody>
<tr>
<td align="center" colspan=2 style="text-align: center"><strong>Reasoning &amp; Knowledge</strong></td>
</tr>
<tr>
<td align="center" style="vertical-align: middle; text-align: center">HMMT (Feb 26)</td>
<td align="center" style="vertical-align: middle; text-align: center">49.0</td>
</tr>
<tr>
<td align="center" style="vertical-align: middle; text-align: center">HMMT (Nov 25)</td>
<td align="center" style="vertical-align: middle; text-align: center">45.0</td>
</tr>
<tr>
<td align="center" style="vertical-align: middle; text-align: center">MMMLU</td>
<td align="center" style="vertical-align: middle; text-align: center">56.0</td>
</tr>
<tr>
<td align="center" style="vertical-align: middle; text-align: center">MMLU-Pro</td>
<td align="center" style="vertical-align: middle; text-align: center">67.0</td>
</tr>
<tr>
<td align="center" style="vertical-align: middle; text-align: center">GPQA Diamond</td>
<td align="center" style="vertical-align: middle; text-align: center">56.0</td>
</tr>
<tr>
<td align="center" colspan=2 style="text-align: center"><strong>Instruction Following &amp; Agents</strong></td>
</tr>
<tr>
<td align="center" style="vertical-align: middle; text-align: center">IFEval</td>
<td align="center" style="vertical-align: middle; text-align: center">64.8</td>
</tr>
<tr>
<td align="center" style="vertical-align: middle; text-align: center">IFBench</td>
<td align="center" style="vertical-align: middle; text-align: center">24.0</td>
</tr>
</tbody>
</table>
</div>

---

## Reasoning

Reasoning (thinking traces) is supported via the chat template. Extended reasoning evaluations are pending the full release.

---
## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load model and tokenizer (trust_remote_code needed for the custom class)
tokenizer = AutoTokenizer.from_pretrained("sKT-Ai-Labs/SPEED-1", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    "sKT-Ai-Labs/SPEED-1",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

# Apply the model's chat template (SPEED uses ### User: / ### Assistant:)
messages = [{"role": "user", "content": "Create a Python CLI tool using docker-py that manages Docker containers. I need a watch command that listens to the Docker events stream and triggers a callback whenever a container dies with a non-zero exit code."}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True,
                                       return_tensors="pt").to("cuda") if torch.cuda.is_available() else \
          tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")

# Generate (give it room to reason before it answers)
outputs = model.generate(**inputs, max_new_tokens=2048, temperature=0.6, top_p=0.9)
result = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(result)
```

> **Note:** SPEED reasons inside <reasoning> tags before answering — keep
> `max_new_tokens` high (≥1024) or the reply is cut off mid-thought.


---

## Citation

```bibtex
@misc{SPEED-1,
  author = {SKT AI LABS, India},
  title = {SPEED-1: Indic-first multilingual Mixture-of-Experts model},
  year = {2026},
  publisher = {SKT AI LABS},
  url = {https://huggingface.co/sKT-Ai-Labs/SPEED-1}
}
```