I Can’t Believe This Model Is Open-Sourced!!!!

If you’ve been following the AI space, you know that the race to build the most powerful, reasoning-capable models has been dominated by big players like OpenAI, Anthropic, and Google. But hold onto your hats, because the game just changed. DeepSeek, a Chinese AI company, has dropped a bombshell: DeepSeek-R1, a fully open-source reasoning model that’s not just competitive with OpenAI’s O1 — it’s blowing minds with its performance. And the best part? It’s completely open-source under the MIT license. Yes, you read that right.

What Is DeepSeek-R1?

DeepSeek-R1 is a reasoning model designed to tackle complex tasks like math, coding, and logical reasoning. It’s part of the DeepSeek family, which recently released DeepSeek-V3, one of the best open-source models out there. But R1 takes things to a whole new level. It’s a thinking model, meaning it uses a process called test-time inference or compute scaling to reason through problems step by step. Think of it as having an internal monologue, where the model debates with itself to arrive at the best possible answer.

What’s even more jaw-dropping is that DeepSeek-R1 isn’t just one model. The company has also released six distilled versions of R1, ranging from 1.5 billion to 70 billion parameters. These smaller models are not only lightweight but also incredibly powerful. For example, the DeepSeek-R1-Distill-Qwen-1.5B model is outperforming GPT-4 on certain benchmarks. Let that sink in: a 1.5 billion parameter model, small enough to run on edge devices, is giving GPT-4 a run for its money.

Why Is This a Big Deal?

1. It’s Fully Open-Source: DeepSeek-R1 is released under the MIT license, which means you can do anything with it — download it, modify it, fine-tune it, or even use it to train new models. This is a stark contrast to OpenAI’s terms of service, which prohibit using their outputs to train other models. DeepSeek is not just open-sourcing the model; they’re encouraging innovation.
2. It’s on Par with OpenAI’s O1: According to benchmarks, DeepSeek-R1 matches OpenAI’s O1 in performance across math, coding, and reasoning tasks. In some cases, it even outperforms O1. For example, the distilled 7 billion parameter model scored 55.5 on the AIME 2024 benchmark, beating models like GPT-3.5 and Claude-3.5.
3. It’s Built Differently: DeepSeek-R1 was trained using large-scale reinforcement learning (RL) without relying on supervised fine-tuning (SFT). This is a groundbreaking approach because it allows the model to explore and learn reasoning patterns on its own, without being spoon-fed labeled data. The result? A model that can self-verify, reflect, and generate long chains of thought to solve complex problems.
4. It’s Fast and Accessible: Unlike some proprietary models that are slow or frequently down, DeepSeek-R1 is lightning-fast and available for free on chat.deepseek.com. You can also access it via their API, which is not only cheaper than OpenAI’s but also has no rate limits.

The Distilled Models Are a Game-Changer

One of the most exciting aspects of DeepSeek-R1 is the distilled models. These smaller versions are fine-tuned using data generated by R1, and they’re proving to be incredibly powerful. For instance:

• The DeepSeek-R1-Distill-Qwen-7B model scored 55.5 on AIME 2024, outperforming GPT-3.5 and Claude-3.5.
• The DeepSeek-R1-Distill-Qwen-32B model is beating OpenAI’s O1-mini on multiple benchmarks.

These distilled models are not just academic curiosities — they’re practical tools that can run on consumer-grade hardware. Imagine having a model that’s as capable as GPT-4 but small enough to run on your laptop or even a Raspberry Pi. That’s the future DeepSeek is building.

Real-World Performance: Mind-Blowing Examples

Let’s talk about what DeepSeek-R1 can actually do. In tests, it has demonstrated an uncanny ability to reason through complex problems, even when the questions are tricky or have no solution.

1. Math Problems: When given a challenging integral problem from the IIT entrance exam (one of the toughest exams in the world), DeepSeek-R1 solved it in 29 seconds, providing the correct answer with step-by-step reasoning. GPT-4, on the other hand, couldn’t solve it.
2. Tricky Questions: When asked a math problem with no solution, DeepSeek-R1 didn’t just give up — it reasoned through the problem, concluded that there was no solution, and explained why. This level of self-awareness is rare in AI models.
3. Chemistry Puzzles: In a chemistry problem where the question contained a subtle error (carbon monoxide instead of carbon dioxide), DeepSeek-R1 identified the mistake, corrected it, and provided the right answer. It even debated with itself, wondering if the user had made a typo.

The Future of AI Just Got More Open

DeepSeek-R1 is more than just a model — it’s a statement. It proves that you don’t need billions of dollars or a massive PR machine to build world-class AI. With the right approach (in this case, reinforcement learning), you can create models that rival the best in the world and make them accessible to everyone.

This release is also a wake-up call for the AI community. While companies like OpenAI and Anthropic have been focused on proprietary models, DeepSeek is keeping the original spirit of open AI alive. They’re not just releasing models; they’re sharing their training secrets, encouraging fine-tuning, and empowering developers to build on their work.

Final Thoughts: Is This the End of Proprietary Models?

Not quite. Proprietary models still have their place, especially for enterprise use cases. But DeepSeek-R1 is a powerful reminder that open-source AI is not just alive — it’s thriving. With models like this, the barrier to entry for AI innovation is lower than ever.

So, if you haven’t already, head over to chat.deepseek.com and give DeepSeek-R1 a try. Whether you’re a developer, researcher, or just an AI enthusiast, this model is worth your attention. And who knows? The next big breakthrough in AI might just come from you, thanks to DeepSeek.

The future of AI is open, and it’s looking brighter than ever.

Have a look into this video running some reasoning problems :