Hi everyone! I’m new here, but I’ve been using Replit Agent a lot lately (like… way too many times a month ). I wanted to suggest something: Would it be possible to add DeepSeek support as the default for Replit Agent?
I’ve heard DeepSeek is super cost-effective while still being powerful, and since many of us rely on the agent daily, this could save credits (and headaches!) for the community. Plus, it might make Replit even more accessible for coders on a budget.
What do you all think? Would love to hear if others are into this idea!
I agree. Loving Replit as a non-coder to develop protoypes. But the agent is simply too expensive to do more serious work and often not that smart. Switching to a lower reasoning model at a lower overall cost could address both?
Moreover, and more importantly, the iteration speed of open-source models will be very fast, which will accelerate the development of the entire world in areas such as software and artificial intelligence.
Like many of the new releases, it is unwise to buy into the hype. OpenAI’s o1, and even o1 Pro, was never the best coding model. Claude maintains that position. Most of the metrics published don’t hold up under scrutiny. Others and I publish comparisons of new models to their claims and different models. These will start coming out in the next few days-it takes time to run these tests thoroughly and effectively.
While I understand your concerns, I believe there are several key points that are being overlooked in your assessment of the new DeepSeek R1 model.
Different Model Architecture: It’s important to recognize that DeepSeek R1, OpenAI’s models (including gpt-4o), and Claude have different architectures, each with unique strengths. The fact that OpenAI and Claude aren’t necessarily the “best” in every use case doesn’t mean other models like DeepSeek R1 can’t offer compelling advantages. The DeepSeek R1’s design is optimized for certain tasks, and that’s a valid choice depending on the specific problem you’re trying to solve.
Open Source Model and Weights: One of the significant advantages of DeepSeek is that it is open-source, including its model weights. This enables the community to iterate, fine-tune, and enhance the model far more rapidly than a closed-source model like OpenAI’s or Claude’s. The open-source nature fosters a collaborative environment where developers can make improvements that align with real-world needs and performance goals. That flexibility can lead to quicker and more targeted updates.
Context Limitations with OpenAI’s Models: You mentioned the O1 and O1 Pro models, but the real challenge with models like OpenAI’s is their smaller context window, which makes it difficult to maintain effective agent conversations or reasoning over long dialogues. DeepSeek R1 addresses this issue with a larger context window, which is crucial for building more coherent, long-term conversations and complex reasoning tasks. For practical applications, having a larger context window isn’t just a feature—it’s a necessity.
Cost Efficiency: Another factor that should be considered is the cost. DeepSeek offers a much more affordable solution, especially for those who need to deploy large-scale applications without breaking the bank. OpenAI’s pricing can be a barrier for many, particularly for businesses looking for flexibility and scalability. DeepSeek’s affordability makes it an attractive choice for those who need performance without the hefty price tag.
I agree that testing and benchmarks are crucial, and I look forward to seeing the comparisons that you and others are working on. But I think dismissing DeepSeek based solely on initial impressions or hype might overlook some of the strengths that it brings to the table, especially in terms of flexibility, cost, and community-driven innovation. #open-source
My point wasn’t that it’s not worth it; my point was to wait until it’s been evaluated by others besides the creators. There are other ‘Open Source’ models that DeepSeek did not benchmark against, such as Llama. There is no 3rd party comparison to Llama published yet.
DeepSeek has also not published their pricing yet, so we really can’t say it’s cheaper. If the goal is cheaper than Sky-T1, which is much cheaper to build, but again, we don’t know about pricing, nor do we know the cost of inferencing if you host it yourself.
We also don’t know that it Is different because OpenAI didn’t publish their technical details in full for all we know it could be a MHA/MOE
Again, my statement was not to do it. It was take a beat, don’t think its great–yet.
In either case, having the opportunity to configure the Agent to use R1, or other models, as a user choice, would be phenomenal. The performance profile and cost tradeoffs might vary subjectively depending on the project in question.
The majority of open-source models you mentioned aren’t even mainstream (including Grok-1). But I like Grok-2 at X frankly!
If you take the time to see what models people are actually using, even just by checking Google Trends, you’ll notice there are only two dominating today: OpenAI and DeepSeek. Gemini and Claude still have some market presence, but they are nowhere near as dominant as OpenAI and DeepSeek.
When we talk about agentic products like Replit Agent, we often consider the underlying large language models (not in order of importance):
Does it have top-tier logic and intelligence?
Is its context capacity sufficient?
Is it affordable enough?
Is it open-source (This allows service providers like Replit to deploy the model on their servers for their customers without any security & privacy risks.)?
However, as mentioned, OpenAI models have limitations with context length, which makes them unsuitable for agentic products since day one.
My hope is that, now that we finally have an open-source model that is both powerful (compared with the most powerful OpenAI series) and affordable (95% less expensive than OpenAI o1), global citizens can come together, pool their resources, and accelerate its iteration to create a prosperity that belongs to everyone!
Initial reviews of DeepSeek, yet to be published, are
The vast majority of their claims hold (rare)
Programming paradigms can be fundamentally different than we are used to seeing in the West - types and structures of functions, specifically in Python and Java
It is holding its own against o1-pro as well
Out-competing Claude in programming, but the caveat see point 2.
Most direct comparisons are being delayed due to Alibaba’s release so it can be included as well
I’m new here, but I’ve been using Replit Agent a lot lately (like… way too many times a month 
). I wanted to suggest something: Would it be possible to add DeepSeek support as the default for Replit Agent?
