Stop using frontier models for your entire OpenClaw stack. It's expensive and wastes your quota. The first step is identifying which use cases CAN be moved to local models. Here's how I do that:

After you showing your full stack a couple of months back, I wondered when this was coming. Gemma4 is too good not to run most of OpenClaw with G4. Still have a frontier as the orchestrator, and will still use codex / Claude for code output, but not for everything.

Yes exactly

switching now.

You look ready

I use frontier for everything, cerebras when speed is absolutely needed. plenty of fronteir plans are just cheap now tbh

They get expensive as you scale up. Plus do you really want to use your quota on simple tasks that can easily be offloaded to local?

Yes because if it can be done locally it can also be done with cerebras with 3,000 tokens per second for little cost. And time is money, your forgetting how valuable time is

Do you see no benefit of running locally? Most tasks I have locally run don’t require super fast speed.

I absolutely do! however! were not there yet. Everytime i use local it's a compromise rn in something i just dont need to compromise. Even embedding i use fireworks qwen and not local, i could EASILY use my 5090 or my 5080 but i dont want my gpus ramped when i can just use cloud

Interesting maybe it’s just a preference thing. I love knowing my GPUs are being put to work

if i could run minimax 2.7 and they ACTUALLY open sourced it, apache licence and all, and it got atleast 60 tokens per second on my 5090. I mean yeah bro id use my gpus for most things. but i cant afford the tech to run the models i want at the speed iwant

moonshot is amazing bang for buck 1mil tokens for $0.60 and further fallback minimax which is 270b (if i’m not mistaken) for 0.30 for one million tokens i dont see justification to run local 3b models

Sorry bro you got like 7575757GB VRAM

i feel like its too much effort to categorize, scope, determine what kind of intelligence is needed for what task . and it feels bad to use "less intelligence" when you could use the better one.

Local models Local models Local models and opensource

I just set up Gemma4 8b with gpt5.4 as my fall back. My main workflow is running system jobs for me. Scan files, extract info, create list. Find files, update info within those files, report results. I also run from jobs. I do this with the web ui. How do i auto model switch?

When you get into exploring this - local models can do A LOT more, some very small than I think most realize.

Which models are the fastest ?

I think you need to give OC a rest and try Thoth if you want unparalleled UX. Just give it a try and you'll see what I am on about - Most AI assistants still expect people to fight with terminals, Docker, and config files. That is exactly where the AI gap begins. Thoth removes that barrier. It is a local first AI assistant with integrated tools, a personal knowledge graph, voice and vision, shell access, browser automation, scheduled tasks, health tracking, and built in messaging channels. It runs everything locally through Ollama and can add cloud models like when you need frontier reasoning. Your data stays on your machine.

Interesting... And a very valid point

based

Delegation per task/function is a huge part of the setup & I unlocked the best experiences once I understood that

The real waste is not using frontier models It’s using them where you don’t know why you’re using them Local vs frontier is less important than whether the system is actually designed to decide

Simpler execution, Openclaw when on the go, and cowork scheduled tasks for everything else. Give them access to the same files. It has worked well and cut my token usage down 80%+

Local models are becoming useable Gemma 4 and Qwen 3.6 have changed the game

This is exactly how we run it. Frontier models on the tasks that actually need reasoning. Smaller models on triage, classification, and formatting. Most people are burning through their best model on work that a local model handles just as well.

This is the way! 🤖

The local openclaw models are better than the borderline ones, which are used very often

Don’t sleep on and

Is your the biggest hurdle the orchestration complexity or just the latent delta between models? We're researching how a 'Shared Mind' layer can act as the glue between these mixed stacks without the manual babysitting. Would love to hear your thoughts if you're interested in testing out our AI agent solution, currently in Alpha stage: