Kimi 2

[jnorthrup]

Kimi 2 is pretty cool.

this is accompanying my desire to run kimi2 free on nvidia
like this

import OpenAI from 'openai';

const openai = new OpenAI({
  apiKey: '$API_KEY_REQUIRED_IF_EXECUTING_OUTSIDE_NGC',
  baseURL: 'https://integrate.api.nvidia.com/v1',
})
 
async function main() {
  const completion = await openai.chat.completions.create({
    model: "moonshotai/kimi-k2-instruct",
    messages: [{"role":"user","content":""}],
    temperature: 0.6,
    top_p: 0.9,
    max_tokens: 4096,
    stream: true
  })
   
  for await (const chunk of completion) {
    process.stdout.write(chunk.choices[0]?.delta?.content || '')
  }
  
}

main();

but not like this

stop-gap resolution

for kimi2 we need to roll back before roles to trust it. mainly for the tools that are for the special roles.

we cannot turn off the annoying TODO tool. this is a fatal gap for kimi2 on roo current

[mrubens]

You can turn off the todo tool in the provider settings

[jnorthrup]

screenshot please, thanks never saw it for looking many times

[jnorthrup]

this is the closest approximation i know

[mrubens]

It's a per-api-profile setting, down in the advanced toggle

[jnorthrup]

up to you how you want your crud scattered around

[mrubens]

Yeah, it seemed to make the most sense associated with the specific model/provider. Hope this helps.

[jnorthrup]

good enough

[jnorthrup]

ok this conversation brings up a lot of drift in roo design. my candid impression.

1. why is that kimi coder very plain direct and clear communication completely ignored ? so what if theres a tool, if the prompt says don't push the big nuke button, and the model can only push the button? that's what we're seeing

2. I want Aider IN VSCODE USING ROO for the crud true story, everyone does when they see gemini rip-- what this means is the model has no roles, no assisted system prompt, and is generally too full of momentum to care as we see above when you stack 3+ long prompts...

3. roles, personalities, avaters, yeah great,I don't care personally; the models are being trained to see past our scaffolding faster than we can write it.