Triava Labs

Sovereign AI infrastructure for independent operators.
Your model. Your voice. Your business.
Manifesto + technical writeup at

[Triava Labs @triavalabs]

Built a self-hosted Gemma 4 node for n8n.
One server.
No cloud API.
Found two Ollama bugs during testing.
Submitted to the Gemma 4 Challenge.
If you're running Gemma 4 on Ollama and seeing empty responses — this is why:
https://t.co/IcSYyiwwBa

[Triava Labs @triavalabs]

Yesterday I posted about a substitution/synthesis split I noticed in Gemma 4.

Ran an 80-call sweep on 26b MoE this morning to test it.

40 substitution calls → 1 unique response.

Bit-for-bit identical text. Same 38 tokens every time. T=0.2. The mode genuinely bypasses sampling.

Cross-confirmed on e4b by @alimafana + Jiwon SEO this week. Joint thread building.

Data: https://t.co/Wa276vJubq

[Triava Labs @triavalabs]

Ran a temperature sweep on Gemma 4 26B MoE last week. Same prompt, three temperatures. Expected variation. Got identical output. Verbatim. Three times.

Then noticed only certain types of prompts behaved that way.

The pattern: canonical-text retrieval ("what's our return policy?") collapses into substitution mode. The model isn't reasoning — it's pulling text from context and rendering it. Synthesis tasks route through reasoning. Temperature affects synthesis. It doesn't affect substitution.

Two operating modes. One model. Same architecture, two paths.

Implication for agentic workflows: answering a policy question and synthesizing a recommendation aren't the same task at different temperatures. They're different tasks.

This finding came from my sovereign Ollama stack. eval_count, prompt_eval_count, raw response bytes per call — telemetry managed APIs don't expose. When you own the inference path, you see the work. When you rent it, you see the bill.

@alimafana ran cross-stack on the managed Gemini side. Flagged the substitution observation as the next piece worth writing. Confirmation incoming.

Full writeup: https://t.co/u6HfIBeuaO

Building sovereign AI infrastructure for operators done renting their intelligence layer

[Triava Labs @triavalabs]

"Fully local" is often a misleading frame.

You can run an agent runtime locally and still call external APIs. You can use a "local" model with a cloud orchestration layer.

True sovereignty means end-to-end control, including the model layer. It's not a checkbox.

[Triava Labs @triavalabs]

What's open: the methodology, the integration node, the deployment recipe.
What's offered: managed deployments for operators who want the system without becoming infrastructure engineers.
The thesis: own your weights, your voice, your data, your model.

[Triava Labs @triavalabs]

A year of building has clarified what Triava Labs is for.

We started exploring sovereignty across multiple domains. The work has converged on one specific bet:

Independent operators should own their AI layer.

[Triava Labs @triavalabs]

The first deployment is the founder's own AI stack — built on her own work before being offered to anyone else.
A full technical writeup, including the open-source integration tooling, ships this month.