It’s time to take the wraps off something exciting that’s been cooking in the lab - DB-1, our framework for supercharging LLM reasoning capabilities.

We wanted to do something new after making our Enhanced Reasoning feature a while ago, and we’re pretty pleased so far, although I would definitely call it a beta product.

Note: While this is shown in the interface as a model selection, under the hood it’s a specific framework wrapped around other models - I don’t want to claim we’re doing something we’re not here, especially as what we ARE doing is pretty unique.

You might have seen buzz around models like o3 and Gemini 2.5, or, if you really pay attention,Sky-T1 that are able to reason by scaling “Test-time compute” - the time it takes to answer you, rather than the time it takes to train the model originally. Well, we’ve been working on our own approach, and it’s… different. In a good way!

What Makes DB-1 Special?

Instead of training a completely new frontier-level AI model (which, let’s be honest, costs a large fortune - Sky-T1’s $450 impressive fine-tune aside), we’ve developed a framework that can wrap around existing LLMs like a cozy superbrain sweater.

We built specific architectural guardrails and reasoning prompts that allow (existing!) models to approach problems in a flexible way. This approach lets LLMs reassess earleir assumptions and challenge reasoning processes in a way that’s a bit more sophisticated than saying “let’s think step by step” (though that trick still holds up well! It’s known as chain of thought, if you haven’t heard of it - and if you haven’t tried using it in your own prompts, give it a shot, you might be surprised how well it works!).

The DatBot Secret Sauce (Well, Some of It)

While I’m not going to explain exactly how DB-1 works, since it’s my own personal approach (You can get some sense of it from just reading the output - DatBot exposes the entire reasoning process), here’s what I can say:

• It’s a framework, not a model - meaning I can use it with different LLMs - I’ve tested with a few, and it works better with some than others.
• Uses advanced prompting techniques (allowing non-reasoning models to reason like o1 or QwQ (which Sky-T1 is modeled after - note we have both QwQ and o1 available in DatBot for you to try).
• Has built-in architectural guardrails to keep reasoning on track.
• Allows for much longer self-consistent output than any other model, including o1 and QwQ responses.
• Works well for creative tasks, unlike traditional reasoning models, due to how it handles the reasoning process.
• Still occasionally gets … lost in thought, and errors out. So treat it a little like a space cadet, and please try not to be upset if it eats some credits now and then - it’s just lost in thought!
• warning - If you use Sherlock as a tone with it… be prepared for some… very long results. Great analysis, just … long. Very long.

The Science Behind It

We’re just one in a long line of researchers who try different versions of scaling test-time compute. Research from DeepMind, published here, as an example, “Scaling LLM Test-Time Compute Optimally”, shows that smart use of test-time (aka answering you) compute can actually be more efficient than training a bigger model. In fact, they found that, optimized, a smaller model could match the performance of a model 14x its size!

Our framework takes this type of insight and applies it in a practical, (mostly) production-ready way that we can then expose to you.

Why This Matters

Look, training massive models is awesome - we wouldn’t be here without it, but not everyone has access to tens of millions of dollars in compute budget (sadly, not even us). We wanted to create something that could democratize advanced reasoning capabilities - something that could work with the models you already have access to, and that we can improve in tandem with the giants of industry, to make their work extend even further for the rest of us.

What’s Next?

We’re still tweaking DB-1, and we’d love to hear from you! How are you using it?

What kinds of tasks have you thrown at it? Did they work? Fail miserably? Chuck your thoughts at [email protected] with experiences or questions.