Documentation › Models

Models

Setup guide· Configurator

Which model plans, which acts, which takes over when a step is hard — and where those models run.

Roles

Console → Configurator

The agent sees the screen through vision models that run inside your instance. On the Configurator page, admins choose a model for each role:

Role	Kind	What it does
Planning	Vision	Interprets the goal and plans the tasks.
Executor	Vision	Decides each step of a run.
Escalation	Vision	Takes over hard steps, recovery and loops.
Primary	Embedding	Embeds application knowledge for search.
Fallback	Embedding	Used when the primary is unavailable.

Built in: qwen3-vl:8b, qwen3-vl:32b and their -instruct versions for vision, plus two computer-use models, hf/beipo-ai/NorthCUA-Nano and hf.co/bartowski/beipo-ai_NorthCUA-Mini-GGUF:Q4_K_M (see measured results); bge-m3 and nomic-embed-text for embeddings. Prefer the -instruct vision models: the thinking ones can spend minutes on one step. A selection applies to new plans and runs; runs already going keep their models.

Model servers

Two kinds of model server, each with a card on the Configurator showing whether it is up:

Ollama — runs most models, including Hugging Face GGUF models named hf.co/<org>/<repo>[:quant].
The Hugging Face server — runs hf/<org>/<repo> models (transformers vision models and sentence-transformers embeddings). It speaks Ollama’s API, so the agent uses its models unchanged.

Both run inside your instance and are set up by Thalamus. Change address on a card tests and saves a new address for every tenant; the agent picks it up within 15 seconds.

Letting ThalamusCUA run the Hugging Face server

Choose an hf/ model when no Hugging Face server answers, and the Configurator offers to start one on the server that runs your instance. ThalamusCUA then keeps it running — across API restarts — until an admin presses Stop on its card (which also has Restart and its Log). The first start installs its environment and can take several minutes. It uses the GPU where it can, keeps two models loaded at most, and unloads one after 30 minutes unused.

Adding and removing models

Add a model by name; it downloads on its server and can be chosen for a role once installed. Removing one you added takes it off the list (an hf/ model can also have its files deleted). A model a role uses, or one still downloading, can’t be removed.

Comparing models: measured results

Configurator note · Console → Documents

A model that has been measured shows its results as soon as you pick it for a role, before you save: a table comparing it with the current executor, what the numbers say, a recommendation, and a link to the full report on the Documents page, with its summary and raw results.

The first evaluation (2026-11-10, a Mac Studio with an Apple M1 Ultra and 64 GB, no CUDA) compared the two NorthCUA computer-use models with the current executor:

Model	How it runs	Calibration	Benchmark	Average step	First load
qwen3-vl:8b-instruct (current executor)	Ollama	8/12	6/10	1.1 s	4.6 s
NorthCUA-Nano 9B	Hugging Face server, bf16	10/12	5/10	15.7 s	60 s
NorthCUA-Mini 35B-A3B, thinking off	Ollama, Q4_K_M	8/12	3/10 (5/10 on a second run)	1.4 s	3.9 s

Calibration asks the model to point at 12 known targets on a 0–1000 grid; 11 of 12 is a pass, and none passed. Benchmark is 10 read-only tasks on the fixture site, so a difference of one or two tasks is within run-to-run noise.
NorthCUA-Nano points most accurately, but its linear-attention layers have fast kernels only for NVIDIA GPUs; on a Mac each step takes about 16 s. It suits instances on CUDA hardware.
NorthCUA-Mini is about 11 times faster than Nano (3B of its parameters are active per step) but no more accurate here, and sometimes answers in pixels instead of the grid. The full model is 70 GB, so a 4-bit build runs on Ollama.
Both were trained on NorthCUA’s own harness and action format; ThalamusCUA asks for its own, which is part of why they score below their published numbers.

Thinking off for NorthCUA. Ollama ignores its “no thinking” flag for NorthCUA’s chat template, so the model would reason until its output budget ran out and answer nothing. ThalamusCUA sends NorthCUA models on Ollama a prompt with thinking already switched off whenever a role runs without thinking — nothing to configure.

The recommendation from that evaluation: keep qwen3-vl:8b-instruct as the executor.

New measurements

Thalamus measures models — calibration on a 12-target page, ten read-only tasks, and the time a step takes — and publishes each evaluation with ThalamusCUA. A new evaluation appears on the Documents page with its report, summary, raw results and the scripts that produced it, and its models get the note on the Configurator. To have a model measured, contact Thalamus support.