DeepSeek-R1: The Free o1 Alternative

How to use the new DeepSeek-R1 model and how it compares to o1.

Hot Takes are occasional timely posts that focus on fast-moving news and releases, in addition to my regular Thursday and Sunday columns.

If Hot Takes aren’t your cup of tea, simply go to your account at www.whytryai.com/account and toggle the “Notifications” settings accordingly:

---

TL;DR

Chinese AI lab DeepSeek just released its newest reasoning model: DeepSeek-R1.

What is it?

DeepSeek-R1 is a free, fully open-sourced reasoning model that performs on par with OpenAI’s o1 across many benchmarks:

Source: DeepSeek

How do you use it?

If you know what you’re doing and want to use DeepSeek-R1 for building apps, fine-tuning, etc., you can access it via DeepSeek API or grab the model on Hugging Face.

If you’re a regular user like me, you can chat with DeepSeek-R1 for free at chat.deepseek.com.

Here’s my quick video walkthrough and a showcase of its capabilities:

Why should you care?

As I see it, the launch of DeepSeek-R1 is a big deal for the average user but also has implications for the industry at large.

User-level implications

With DeepSeek-R1, everyone now has access to a reasoning model that:

1. Performs on par with o1 (almost): DeepSeek-R1 has similar scores on many benchmarks, although observers like AI Explained suggest it has its blind spots.

2. Is free to use and cheap to build with: DeepSeek-R1 is free for regular chats, and its API pricing per 1M output tokens is almost 30 times cheaper¹ than o1.

3. Provides better insight into its thinking process: DeepSeek-R1 offers a more detailed and transparent picture of its inner monologue than the distilled summary from o1 (see above video). This gives users a better chance to trace the reasoning behind the answer, identify where it goes off the rails, and perhaps steer the model better in subsequent requests.

4. Can render the resulting code: As I’ve shown, DeepSeek-R1 lets you test the apps it creates directly in the chat interface (a la Claude Artifacts). This is helpful for quick back-and-forth iterations.

5. Is fully open-sourced: Developers can fine-tune the model, distill it, and otherwise access the underlying code.

Broader implications

DeepSeek has shown that it’s possible to quickly develop and open-source an o1-level reasoning model² while making it less expensive and more transparent than OpenAI’s proprietary ones.

We’re only three weeks into January and already have a capable open-source competitor to OpenAI’s o1. I’m starting to think this two-week-old prediction of mine might’ve been too conservative:

By the end of 2025, reasoning models from at least three other players will perform on par with or better than OpenAI’s o3.

🫵 Over to you…

What’s your take? Are we about to see an avalanche of reasoning models from other firms? Have I overlooked some important implications?

If you’ve had the chance to test DeepSeek-R1 and compare it to o1 for your tasks, I’d love to hear what you think!

Leave a comment or drop me a line at whytryai@substack.com.

1. DeepSeek-R1 is $2.19 per million output tokens (here). OpenAI o1 is $60 per million output tokens (here).

2. On the other hand, I have yet to see any serious discussion of what these condensed timelines mean for safety testing, alignment, etc.