Open models,
where they pay off.

Open-weight models (often called open source) can run anywhere, be fine-tuned on your data, and cost a fraction per token. Humonix uses them inside your Brain for the work they do well, and keeps frontier models for the work that needs them.

FRONTIER · HARD WORKClaudeGPTGeminiGrokOPEN WEIGHTS · BULK WORKLlamaDeepSeekQwenMistralGemmaYOUR BRAIN · YOUR DATA STAYS HERE

What “open” gets you

Run it where you want
On a managed host, in your own cloud account, in a Canadian region, or on your own servers. The model goes to the data, not the other way around.
Shape it to your business
Fine-tune on your own product data, tickets and documents so it speaks your language and knows your catalogue.
Pay for what you use
Per token on a shared host, or a fixed monthly price for dedicated capacity when volume is steady.
No lock-in
The weights are published. If a provider changes its terms, the model and your fine-tune move with you.

When we use them, and when we don’t

Good fit
  • Bulk work: tagging, extraction, scoring, summaries over thousands of records
  • Data that must stay in a region or on your own infrastructure
  • A model fine-tuned on your catalogue, tone or support history
  • Steady, high-volume jobs where cost per token decides
Not the right tool
  • The hardest reasoning and long, multi-step agent work
  • Customer-facing writing where quality is the product
  • Low volume, where setup outweighs the savings
  • Anything a smaller, cheaper frontier model already does well

The families we work with

Meta
Llama
General-purpose family in many sizes; a common base for fine-tunes.
DeepSeek
DeepSeek
Strong reasoning and code for the price.
Alibaba
Qwen
Many sizes, many languages, good at structured output.
Mistral AI
Mistral
Strong French, with open-weight models in several sizes.
Google
Gemma
Small, efficient models that run on modest hardware.

Where they run

Managed host
Per-token pricing on providers that serve open models. Fastest to start; nothing to maintain.
Your cloud
Deployed in your own AWS, Azure or Google Cloud account, in the region you choose, including Canada.
Dedicated capacity
Reserved GPUs for steady volume: a fixed monthly price instead of per-token billing.
On your premises
For data that cannot leave the building. Smaller models run on modest hardware.

What moving bulk work to open weights saves

Your monthly volume and the share of it that is bulk work. Prices default to public list prices; change them to yours.

Saved per month$936
Cost today$1,800
Cost with open weights on bulk work$864
Saved per year$11,232
%52%

Defaults: Claude Sonnet 5.5 and an example open-weight model on a managed host, blended at 80% input and 20% output tokens, list prices as of October 2026. Savings depend on the job: test quality on your own data first, which is part of every Humonix setup.

Questions

Are open models as good as Claude or GPT?

Not across the board. On narrow, repetitive jobs they are often close enough, and much cheaper. We test them on your own examples before switching any job.

Is it really open source?

Many are “open-weight”: the weights are published under a licence that may limit some uses. We check each licence against how you plan to use it.

Can it run in Canada?

Yes. In your own cloud account in a Canadian region, or on dedicated capacity there. Managed hosts vary by region, so we confirm it in writing before any client data goes there.

Do I have to choose one model?

No. Your Brain uses several: frontier models for hard work, open-weight models for bulk work, and fast decision models for scoring and routing.

Find the jobs worth moving

We look at your volumes and data rules and tell you which jobs, if any, should move to open weights.

Book a call