Download
← All posts
Tutorials Jun 11, 2026

LLM Profiles - how Skales helps every model use tools properly

Tool-calling reliability varies wildly between models. LLM Profiles tune tool budget, prompts and sampling per model family - so even a small local model stops fumbling. Here is how it works.

There is a dirty secret in the agent world: the same agent that works flawlessly with a frontier model falls apart on a smaller one. Not because the small model is stupid - but because tool-call reliability and the right sampling parameters vary a lot by model, and one global setting can never fit them all.

That is the problem LLM Profiles solve. Since v11.3 they ship inside Skales, and the profile catalog is open source at github.com/gonemedia/llm-profiles.

What a profile actually does

A profile is a small, reviewable JSON file that matches a provider/model pair and tunes exactly four things:

Which models are covered

Built-in profiles exist for DeepSeek, Qwen, GLM, MiniMax, Kimi, Mistral, Gemma, Llama, Phi, GPT-OSS and a generic profile for models of 9B parameters or less. Frontier models - Claude, GPT, Gemini - intentionally have no profile and run unchanged. They do not need help.

How to use it

Settings → AI Providers → LLM Profiles. It is opt-in: flip it on and the best-matching profile applies automatically (most specific match wins, your imported profiles outrank the built-ins). You can import your own via file, JSON paste or a GitHub URL - and contribute profiles back via pull request.

If you run local models through Ollama or LM Studio, this is the single biggest quality lever in Skales. A guardrail also warns you when a very small model is paired with too many tools.

Try it with your favorite local model - Skales is free for personal use.

New here? Skales is the private AI agent you install like an app - free for personal use, any model, everything stays on your machine.

Download Skales

More from the blog

Updates Sep 6, 2026 Skales v12.9.26 - Grip Grip is the release where Skales stops when you say stop, and tells you what it costs before it spends. Under every reply sits the price of the session, read from the provider or from the price table, with the sub-agents' bill included; a ceiling you set in Goals stops a run before the next paid step and asks you three things - continue, take a cheaper model, or stop - in the chat, in the Code window, in Iris and on the phone. A typed "stop" is honoured before the first request leaves the machine, and a stopped run says so in its own bubble instead of pretending it finished. Credentials are locked away from the file, shell and search tools; a bulk action or a publish asks first, in every mode, and a yes never covers more than the card names. Updates Sep 4, 2026 Skales v12.9.25 - Backbone Backbone is the release where Skales got one engine. A chat you type, a task you scheduled, a job on a timer, a plugin, a webhook, a Discord message, a Telegram or WhatsApp turn and the desktop Buddy used to run on different loops with different rules; now they all run on the same one, with the same guards, the same budget and the same record of what happened. Close Skales in the middle of a run, update it, let it crash: the run comes back as an unfinished run, says how far it got, and one Continue picks it up at that step. Ask for something big and Skales can hand parts of it to sub-agents, each in its own checkout of your repository, and collect the results. Under every reply sits a context meter that names what filled the window: the system prompt, the tool definitions, your skills, your memory and the conversation itself. Updates Sep 2, 2026 Skales v12.9.21 - Cockpit 2.0 Cockpit 2.0 is named for the screen it replaces. What used to be two stacked tab bars, a second title and a row of icon chips is now one board: everything that stands, in four columns, with goals, your own tasks, Autopilot tasks and schedules as cards you can filter, drag and act on. One head, one action. The phone got the same board in 2.9.21, and a computer on an older version still fills it from what it can tell rather than showing an empty screen.

All posts →