I built the fastest coding agent on earth. 14,500 tokens per second. Seriously. It can create thousands of websites before Paul Jankura Fable can even make *1*. I'm using the Claude R Perrin harness and powering it through a custom proxy with Taalas Inc. silicon models.

A little demo of the speed (not quality- it's a 3b param model after all)

Done in 0s.

Websearch to learn how to run a subprocess. 1s.

First some caveats: While yes, it really is Claude Code running at 14,5000, it's not an @anthorpi model. It's Llama 3b. The biggest challange was @taalas_inc doesn't offer a public API, but they do offer a demo chat bot: (I think you know what happens next)

Taking a page out of the "ChippotleAI" project I thought, "Chatjimmy looks like a public API to me!" So I built a local proxy that uses to get access to their hardware silicon model. And Wow. It is insanely fast.

The problem is though: Llama 3b is horrible for an agentic coding harness. And Taalas's chatjimmy API doesn't offer any sort of tool calling. So I had to build all of that.

I used the same logic behind @karpathy Autoresearch to get a Claude Opus workflow to find me: - The best system prompt to force Llama 3b to become an agentic programmer with tool calling - The best combination of Claude tools to expose to the model to keep it focused.

I also had to: - repair broken tool calls - block dangerous tool calls (First thing It did was try to 'sudo rm -rf' my whole system. Claude said it was 'just a hallucination'- I'm not so sure...) - merge multirounds of inference to get a solid response

The final result is a thing of beauty. Try it for yourself while the @taalas_inc team is still asleep, by morning they'll probably block and it won't work anymore. But it will have been a beautiful glimpse of the future while it lasted.

@Claude @taalas_inc @taalas_inc this is my way of trying to impress you to ask for api access. Pretty please?

@Anthropic @Claude @taalas_inc Anthropic is sleeping, Taalas is still sleeping, only you out here cooking at 14.5k tokens/s Absolute legend 😂

@Anthropic @Claude @taalas_inc 14,500 tokens per second and you said seriously like we might not believe you lmao

@Anthropic @Claude @taalas_inc mad respect, that speed is insane

@Anthropic @Claude @taalas_inc is this expensive to run since you are using more tokens per sec?

@Anthropic @Claude @taalas_inc Noooooope. It’s free. (Because the chatbot is free)

@Anthropic @Claude @taalas_inc That’s sick. Waiting for reactions from @taalas_inc 🤭 Ca: EDjvJPVK4Ue4Rkh83dJrc7whzBqQBcu5EfhicmuCpump

@Anthropic @Claude @taalas_inc So basically, AI slop built faster?

@Anthropic @Claude @taalas_inc Much, much, faster.

@Anthropic @Claude @taalas_inc Someone from the bankr community launched a coin with fees directed to you. you can use the fees towards building the platform or even integrate a token into your tech CA：0x356fd44dfbe8a7bc4cec2ca9f014dd038dacaba3 this thread also explains a bit more.

Hey Dennison, Your Claudberghini project is genuinely impressive — the quality is next level. Hitting 14.5k TS already is wild, especially knowing someone dropped the token with fees routed straight to you. Contract: EDjvJPVK4Ue4Rkh83dJrc7whzBqQBcu5EfhicmuCpump Whenever you have a moment, claim the bag and let me know your thoughts. If you’re down, we could actually build something meaningful with this instead of just farming it. What do you say?

@Anthropic @Claude @taalas_inc that's very cool sir

@Anthropic @Claude @taalas_inc I sent you a dm would appreciate a reply

@Anthropic @Claude @taalas_inc Actually, I wonder if I should use this to power @typeROID

@Anthropic @Claude @taalas_inc @typeROID Well no, because @taalas_inc will probably take it away soon. I'm being sneaky with my proxy to get access to this model.

@Anthropic @Claude @taalas_inc How do the @taalas_inc silicon models change the proxy latency compared to standard setups? ⚡

@Anthropic @Claude @taalas_inc would you be open to claiming fees on a coin on @pumpfun? ive generated creators thousands of dollars because of their insane tech let me know, I can explain in the DMs

@Anthropic @Claude @taalas_inc I have wondered if 1. This work work 2. If anyone would try it 3. If Taalas would notice and allow it Very very cool. I am very excited for the possibilities around this sort of approach.

@Anthropic @Claude @taalas_inc 又来刷数据了是吗？

@Anthropic @Claude @taalas_inc Thoughts about receiving royalties for your product?

@Anthropic @Claude @taalas_inc 14,500 tokens per second through a custom silicon proxy is the kind of number that makes you want to see what it does with an hour of unsupervised runtime

@Anthropic @Claude @taalas_inc 14 500 tokens per second is mind-blowing

@Anthropic @Claude @taalas_inc Hey we launched a bankr for Claudberghini. If you're not sure how it works you can read up on it here. Just a way for us to help fund it if you wanted some form of donations

@Anthropic @Claude @taalas_inc CA: EDjvJPVK4Ue4Rkh83dJrc7whzBqQBcu5EfhicmuCpump Claim the bag whenever you can and let me know your thoughts. If you’re up for it, I’d love to build something meaningful together