Introducing DeepThought-8B: A small, capable reasoning model
Our first release: a small reasoning model built on LLaMA-3.1 8B that breaks down its thinking into transparent, structured steps.
Ruliad team2 min read
Today we're releasing DeepThought-8B, a reasoning model built on LLaMA-3.1 8B. This release is our first step toward making AI reasoning more transparent and controllable, and a test of how far an efficient 8B model can go when it reasons step by step. DeepThought-8B unlocks test-time compute scaling during inference for all, taking as many reasoning steps as needed to solve complex problems.
We're excited to make DeepThought-8B available today, with controls that let you shape the way the model reasons.
What makes DeepThought different?
DeepThought-8B approaches tasks differently from traditional language models. Given a problem, it breaks down its thinking into clear steps until it reaches a conclusion. Here's what this looks like in practice:
{
"step": 1,
"type": "problem_understanding",
"thought": "The user is asking how many Rs there are in the word 'strawberry'"
}Each step is documented, making it easier to understand how the model arrives at its answers.
Key Features
- Transparent Reasoning: The model shows its work, step by step.
- Programmable Approach: You can guide how the model reasons without retraining it.
- Test-time Compute: The model can take as many steps as needed to solve a problem.
- Runs Locally: At 8B parameters, DeepThought runs on consumer GPUs with 16GB+ VRAM, so you can run it on your own machine.
- Structured Output: Consistent JSON-formatted reasoning chains for easy integration.
Performance
While we're excited about DeepThought-8B's capabilities, we believe in transparency and community-driven evaluation. Our internal testing shows promising results across reasoning, math and coding benchmarks, but we encourage you to test these capabilities yourself.
Overall benchmarks
| Model | All | Math | Reasoning | Coding | IF |
|---|---|---|---|---|---|
| DeepThought-8B | 55.91 | 55.68 | 69.34 | 25.61 | 84.03 |
| Qwen-2-72B-Ins | 56.24 | 72.21 | 57.17 | 31.23 | 66.39 |
| Qwen-2.5-72B-Ins | 52.53 | 69.30 | 51.12 | 23.35 | 69.15 |
| Llama-3.1-70B-Ins | 47.34 | 57.05 | 39.13 | 24.58 | 62.34 |
| Llama-3.1-8B-Ins | 45.90 | 53.51 | 36.78 | 23.07 | 61.86 |
| Claude-3.5-Sonnet | 69.17 | 70.48 | 93.18 | 39.91 | 88.82 |
| GPT-4o | 70.99 | 74.78 | 89.22 | 46.75 | 88.37 |
| o1-mini | 71.70 | 91.04 | 91.67 | 42.89 | 85.67 |
Insight
DeepThought-8B, at just 8B parameters, achieves reasoning and instruction-following scores competitive with models 9x its size, and outperforms Llama-3.1-70B on both reasoning and IF benchmarks.
Some early findings
- Strong performance in step-by-step problem-solving
- Competitive results on coding and mathematical tasks
- Reliable instruction following with transparent reasoning chains
- Performance scales with test-time compute, allowing deeper reasoning on complex tasks
Limitations
Like all models, DeepThought-8B has its limitations. We're actively working on:
- Improving mathematical reasoning for complex problems
- Enhancing long-context processing
- Increasing competence in edge cases
Rather than hyping benchmark scores that might not reflect real-world usage, we invite you to:
- Test the model in your specific use cases
- Share your findings with our community
- Help us identify areas for improvement
You can report your findings and discuss the model's performance by tagging or DMing @ruliad on X. You can also send us an email at feedback@ruliad.co.
Download
Download DeepThought-8B-Llama-v0.01-alpha now on our Hugging Face repo.
What's next? This is just the beginning. We'll be regularly updating the model based on your feedback and our ongoing research.
Cite this
@misc{ruliad2024introducingdeept,
title = {Introducing DeepThought-8B: A small, capable reasoning model},
author = {Ruliad team},
year = {2024},
publisher = {Ruliad},
url = {https://ruliad.co/research/introducing-deepthought-8b}
}Working on something related? We'd like to hear from you.
Get in touch →