Terp

Be kind to others, everyone deserves it

- Vibes @potionalpha

- Sponsor @spumpio

[Terp @OnlyTerp]

ULTRACODE-SHIM IS NOW LIVE 🔥

You can now run ANY model in UltraCode

I built a github repo to make this really easy for you, Just send your agent there and let him COOK

You deserve the flexibility to use LOCAL models & cost efficient models. So I made that happen for you 🫶

[Terp @OnlyTerp]

Composer 2.5 in ultracode is INSANE 🤯

[Terp @OnlyTerp]

Get ready for the newyork tiktoks of people going absolutely crazy in the streets after that game. Wow

[Terp @OnlyTerp]

@HermesAgentTips I use mimo a ridiculous amount. Super dependable model tbh. Would say he is the smartest for me behind claude and gpt

[Terp @OnlyTerp]

@TheAhmadOsman i have a single 5090 and a 4090 & I feel like it works really well for me right now. I'd really like to get more optimization support for them tho, they have alot more power then i can squeeze out

[Terp @OnlyTerp]

@Ominousind haha

[Terp @OnlyTerp]

@TOMAW135 forgot to add a space kek

[Terp @OnlyTerp]

Here's some benchmarks to see how the new Gemma 4 12b stacks up. Ran on my Linux Pc with a RTX 4090

Great at coding, outstanding actually
Stuggles with toolcalls
Solid speed
Not the best with following instructions

(the 31b model in the comparison testis gemma opus finetune)

[Google Gemma @googlegemma]

Meet Gemma 4 12B!

A unified, encoder-free multimodal model designed to bring high-performance intelligence directly to your laptop, and released under an Apache 2.0 license.

Bridging the gap between edge efficiency and advanced reasoning. Here is what’s new with Gemma 4 12B: 👇

[Terp @OnlyTerp]

@keylimesoda https://t.co/nTGtWSke1y by @outsource_

[Terp @OnlyTerp]

@dxmmhet and i was using devin cli to run it with mimo v2.5pro orchestrating to run the tests and stuff

[Terp @OnlyTerp]

@dxmmhet i used https://t.co/nTGtWSke1y by @outsource_

[Terp @OnlyTerp]

@jackowhf mimo but like its so big can u count that since we cant run it local haha, ur best bet for re is get the $6 mimo plan and the $20 minimax plan and have mimi v2.5pro orchistrate and use a ton of minimax m3 subagents. ud get a ton of usage for that and it would serve u well

[Terp @OnlyTerp]

@TotallyNotParth LOOOL old pc has 64 gb ddr5 and ryzen 9 and 4tb nvme KEK it was for my office

[Terp @OnlyTerp]

@Ominousind https://t.co/oAxxcFd7yF we are aligned in our research

[Terp @OnlyTerp]

Here's some benchmarks to see how the new Gemma 4 12b stacks up. Ran on my Linux Pc with a RTX 4090

Great at coding, outstanding actually
Stuggles with toolcalls
Solid speed
Not the best with following instructions

(the 31b model in the comparison testis gemma opus finetune)

[Terp @OnlyTerp]

100% local swarm

1 RTX 5090

[Terp @OnlyTerp]

Caduceus now supports auto local routing

What this means, caduceus understands what local models you use & will automatically choose the best local models for the job

Models will easy unload & swap in vram when needed on demand

Try it today <3

[Terp @OnlyTerp]

Just hit 50k followers! 🥹

Wanted to say, I've been here for awhile & I couldn't be more grateful to be apart of this community with you all

I never would have imagined to have so many people in my life, I'm honored and I dont take it for granted at all

Thank you so much 🫶

[Terp @OnlyTerp]

@TheAhmadOsman Its been a nightmare but ive been fixing every issue i have with them 1 at a time and posting on github for free to help others https://t.co/dATjP6Z0uI

[Terp @OnlyTerp]

It took me a lot of time, billions of tokens, hairloss, stress, loss of sleep

BUT I DID IT, I FIXED RUNNING LLMS ON WINDOWS!

If you've been using Windows and WSL and it's been giving you tons of headaches, you can finally relax

I made a guide to make it as stable as Linux 🫡

[Terp @OnlyTerp]

@the_dt1 @NousResearch It uses them. The todo list is my personal choice for most tasks as i find it gets the job fully finished the proper way instead of wasting my time. The workflow is a honest implementation of ultracode dynamic workflows. It is fully open source. Fully customizable. Make it yours!

[Terp @OnlyTerp]

I built Caduceus for Hermes Desktop @NousResearch

Two snakes, one staff
Caduceus runs parallel AI agents in perfect sync, on any model you already have

It sizes up every task and picks the path: • Easy → answered instantly • Hard → a live to-do plan + agents fanned out in parallel, verified before it calls it done

Caduceus has a UI right inside Hermes so you watch every agent build your project in real time

Full demo 👇

[Terp @OnlyTerp]

@TiQTiQB00M @NousResearch awesome, so since its dynamic it will actually build a workflow for the task or you can give it a workflow and it will use this system to ececute it. I'm working on a bunch of things to speed it up without sacrificing quality, next few updates will make it much faster

[Terp @OnlyTerp]

@UsmanIdris16845 myb king, all the info is on the github fully open source <3

[Terp @OnlyTerp]

Caduceus now supports auto local routing

What this means, caduceus understands what local models you use & will automatically choose the best local models for the job

Models will easy unload & swap in vram when needed on demand

Try it today <3

[Terp @OnlyTerp]

I built Caduceus for Hermes Desktop @NousResearch

Two snakes, one staff
Caduceus runs parallel AI agents in perfect sync, on any model you already have

It sizes up every task and picks the path: • Easy → answered instantly • Hard → a live to-do plan + agents fanned out in parallel, verified before it calls it done

Caduceus has a UI right inside Hermes so you watch every agent build your project in real time

Full demo 👇

[Terp @OnlyTerp]

@jackowhf @NousResearch LOOL this poor chair