This is why you’re not using open models!

Those eyes, the eyes of someone that has looked deep into code and the code looked back at him. Not even the abyss could believe the $1 GO plan.

Hahaha!! Eyes of someone who went out for a run after two months of dev work hhaha.

Understandable, I recently did some harness engineering to make Gemma 4B model at par with Frontier models for credit card spend analysis. What changed the game was adding proper rollback for long horizon performance. These are super powerful if harness is potent enough.

Aha tell me more

Working on a blog post to document this. Will send it across soon.

I've tried using DeepSeek in command code, I must say it's really the best

Thanks for sharing. Excited to share what we’re cooking 🧑‍🍳

Yesterday, I tried using DeepSeek V4 Pro through Command Code, and it built a very complex feature in one go. It added almost everything, around 90% of the implementation. The only thing I noticed is that it still struggle to identify very complex edge cases on its own.

Goad to hear that. Complex edge case is a model capability problem bot the harness problem. When in doubt try Kimi K2.6.

Incredible work, Ahmad. Thank you 🫡

Thannnk you!!

Already subscribed yesterday to go plan , will share feedback shortly . It will be good to have agent teams in command code to run parallel agents as i am building a complex app with lot of features

Sub agents are there. Run /agent.

@CommandCodeAI Would this still be applicable if I use the command code API in Opencode, for example

@CommandCodeAI Not at all. API has none of this. I recommend you use Command Code to get all this. We need to control the agent loop to help.

@CommandCodeAI Not a big fan of CLI environment and it seems command code doesn’t have a macOS app

i think we should to using open models

Deep seek is really making my life hard

why

Command Code’s $1 GO plan is silently reshaping how we think about model efficiency and cost. The infrastructure moves like it’s built for what comes next, not what’s already here.

Thanks 🙏

Unsloth studio heals toolcalls b4 harness sees it for freeèeeeeeeee

The pricing is impressive, but for serious projects the bigger question is reliability: terminal stability, Windows/PowerShell support, permissions, and predictable file edits. Cheap credits are nice, but accuracy and workflow stability matter more.

Thanks I will stay with Claude code 🌚

For sure. This is what Claude Code does to open models.

i see many peps saying ds v4 flash (xhigh) is noticble better to than ds v4 pro. What u think bossman?

Flash is unbelievably good for the price and size. I use it a lot.

I'm convinced. Will try it out.

I'm not sure. Opencode Zen has pretty much the same performance

Thousands of developers have shared otherwise. You don’t have to take my word or user Command for it. My eng deep dive has everything you need to fix it.

Are these stats public? Because in the opencode Zen version I haven't run into a single Instance of tool calls failing or going off the rails. Would be great if you can share the stats if they are public and benchmarks you guys did.

And no. Tweets don't count. I'm more interested in measurable differences.

You can measure both and report back. I don’t honestly have anything more to share beyond what I discovered and fixed and what is now confirmed by so many people very publicly.

Converted many many naysayers once they tried.

Bro when android users can get this feature

I have had zero issues with deepseek-v4-pro

Great!! Have you tried it with Command Code?

i need GLM 5.1

it's available in our models list.

@grok can you add subs

setup cost is the real blocker. most people aren't gonna spin up an ollama instance just to try it. that gap needs to close

@grok summarize this clip

four fixes and it beats opus? prove it

Try it yourself buddy

on it

make the same fixes for Minimax M3, its really good model and I like it alot so far. would be really good to get this on!

M3 is really good. We have been collecting errors on that and repairs going live soon as this week.

Glad to hear that mate! Would be exciting to see how good it would be

: rm -rf /*

i’d bet my casino on command code that harness hit different

Plenty of folks skip open models because performance and reliability still lag behind closed options like GPT-4, especially for complex tasks.

Not the case anymore especially not with @CommandCodeAI

What is the metric for when you say “outperform” in this case?

internal benchmarks of hard SWE questions, all solved, where dsv4 pro was cheaper, and faster.

@grok ne diyor

Today's underdog is tomorrow's default choice. We've seen this movie before. 👏🔥

Looking forward to more 👀

Habibi.. In the cmd install file theres a lot of your name there. Easter egg? Haha. Anyways, good product.

What?

cool

Type "I want" and let your courage rise