Playing around with the new DeepSeek-V3, and all the latest changes to Bolt.diy (notably the ability for it to troubleshoot errors near automatically), make for a compelling case to use Bolt.diy over other options like Cline (probably the most compelling competitor IMO). And even without MCP’s or tools, which seems absurd.
Using the same model, Cline not only failed to Build a todo app in React using Tailwind but used a ridiculous amount of tokens. I tried several prompt samples and saw basically the same trend each time.
Bolt.diy with the optimized prompt option enabled, not only succeeded near one-shot (it needed to fix an error and prompted me to do so), it used “Tokens: 4686 (prompt: 3542, completion: 1144)”… two orders of magnitude fewer tokens than Cline to do the same task (if I had let it continue no doubt)…
Keep in mind, this is me only letting Cline do it’s thing and not trying to work through the failed task any further. At this rate, I can only imagine how many tokens it would have used and how much frustration I would have had to endure.
4686 (Bolt.diy - Successful) vs 302K (Cline - Failed) Tokens.
And I know this is a very subjective test, it’s literally the first day DeepSeek-V3 is available, and there’s no apples-to-apples comparison here (especially being that Cline is optimized for Claude)… but the results surprised me, so I thought I should share.
And of course the price makes this model very compelling as well.
I’m excited about the recent updates and have been trying out the new DeepSeek-V3 Release in Bolt.diy; However, I’ve encountered a persistent issue that’s preventing me from using it correctly.
Every time I try to use DeepSeek-V3, I receive the following error message:
"There was an error processing your request: An error occurred."
I’ve tried a few things to troubleshoot this, including:
Selecting different models: As you can see in the attached image, I’ve tried both “Deepseek-Chat” and “Deepseek-Coder,” but the error persists in both options.
I’m wondering if anyone else has experienced this same issue or if anyone from the community or the Bolt team could offer some guidance. Any suggestions or advice would be greatly appreciated.
I’m eager to utilize the new capabilities of DeepSeek-V3, and I believe it could be a valuable tool for all of us. Thank you in advance for your time and help!
i got 70 million tokens for 20 bucks, threw it into diy last night, very fast and good and used ~39k tokens to throw a react/node repo together. so far pretty pret-ty good and deepseek tokens are on sale at v2 prices til feb
edit my bad its the usage on sale not the tokens. so got 5 weeks to smash it hard before it goes 4x
bunch more dicking around (i am noob+ level keen to learn nocoder)
deepseek is dope i got bogged down using windsurf free trial and bolt new free trials at getting api calls to my db and a succesful 0auth call to an external api.
after using 3 cents of my deepseek credit i got it all working. i am sure there are issues w the 64k context window for heavy duty stuff but for the price the speed and the results i am honestly blown away. just my littleguy perspective. its so disjustingly cheap , just throw 5 or 10 bucks at it and give it a whirl imo
last update - once i hit around 38k-40k token prompt size it started hallucinating. same issue as all the other ai coders but i actually got a bit further than the rest of them.
I’ve seen someone else have the same issue using Deepseek directly. I’ve only tried it with Openrouter (as I don’t really want more accounts to manage) and have had no issues. As another note, their site has seemed to be down a few times which probably makes sense due to the amount of traffic they have probably been getting.
Me too and their free model sucks. And there are no other options, lol.
And I’ve noticed the same hallucinations after using it for a while… not sure what it is but seems to be related to “memory”. Deleting the chat history (export and save if you’d like) generally solves it for me and I use the optimized prompts in Bolt.diy settings.
