Docs · Models

The model catalog

A small, curated set chosen to span the cards people actually own, from a compact 8 GB starter up to a flagship. Plus how the weights reach a machine, and how Nodeau knows they're the right ones.

The curated catalog#

nodeau model list          # what you can run, and what's downloaded
nodeau model list --all    # including older ids that newer models replaced

Eight recommended models, arranged on a four-rung hardware ladder. Each one says which rung it's designed for. A rung is guidance: the real decision is made by the fit check, on your card, at the moment you ask, and a model can fit a smaller card than its label suggests.

Every model runs on Linux machines with NVIDIA GPUs. On a Mac, which runs standalone, Nodeau serves the chat models. The tasks beyond chat are covered in embeddings, reranking, tools and vision.

Here's the list on a machine that has a few of them already:

MODEL                      ROLE                   GPU     SIZE        STATUS          QUALIFICATION
bge-reranker-v2-m3-q8_0    Reranker               8 GB    606.23 MiB  downloaded      Nodeau curated
gemma-4-12b-qat-q4-0       Mainstream multimodal  12 GB   6.50 GiB    downloaded      Nodeau curated
gemma-4-e4b-qat-q4-0       Compact multimodal     8 GB    4.80 GiB    not downloaded  Nodeau curated
gpt-oss-20b-mxfp4          Large efficient MoE    16 GB   11.28 GiB   downloaded      Nodeau curated
qwen3-embedding-0.6b-q8_0  Embedding              8 GB    609.54 MiB  downloaded      Nodeau curated
qwen3.5-4b-q4km            Compact general        8 GB    2.55 GiB    not downloaded  Nodeau curated
qwen3.5-9b-q4km            Mainstream general     12 GB   5.29 GiB    not downloaded  Nodeau curated
qwen3.8-27b-q4km           Flagship general       24 GB+  17.67 GiB   not downloaded  Nodeau curated

nodeau model list on your own machine is the authoritative list for the build you have. Older ids stay valid for good and keep naming exactly the same bytes. When a newer model takes an older one's place, the older id is marked deprecated, and it still installs and runs.

What a model can do#

A model has a task, which is the kind of work a workload is started for, and a set of capabilities, which are what it has been shown to do.

A workload answers its own task's routes, which keeps every reply the kind you asked for. See which routes an endpoint answers.

nodeau run <model>                      # the task is inferred when a model does only one
nodeau run <model> --task embed         # otherwise, say which

The catalog lists each model for exactly the tasks and capabilities it has shown, and embeddings, reranking, tools and vision has a working example of each.

Looking at a model before you download it#

nodeau model info qwen3.5-4b-q4km

Qwen3.5-4B Q4_K_M
  id            qwen3.5-4b-q4km
  family        Qwen3.5
  parameters    4B
  quantization  Q4_K_M

Licence
  Apache-2.0
  https://huggingface.co/Qwen/Qwen3.5-4B/blob/main/LICENSE

Where the bytes come from
  https://huggingface.co/unsloth/Qwen3.5-4B-GGUF/resolve/e87f1764.../Qwen3.5-4B-Q4_K_M.gguf
  repository    unsloth/Qwen3.5-4B-GGUF
  revision      e87f176479d0855a907a41277aca2f8ee7a09523   (pinned; not a moving branch)
  size          2.55 GiB
  sha256        00fe7986ff5f6b463e62455821146049db6f9313603938a70800d1fb69ef11a4

  Nodeau downloads this from its publisher and never republishes it.
  No account or token is needed.

Capabilities
  chat                yes
  structured output   yes
  tools               yes
  ...

Measured configurations
  ENGINE     CONTEXT  PARALLEL  GPU                         PEAK VRAM  SCOPE    CONFIDENCE
  llama.cpp  4096     1         NVIDIA GeForce RTX 5060 Ti  3,000 MiB  process  measured

Everything you'd want to know before downloading several gigabytes: where the bytes come from, what licence they carry, what the model can do, and which serving configurations have actually been measured on which hardware.

Downloading#

nodeau model install qwen3.5-4b-q4km

Nothing is downloaded until you say yes. The size, licence and source come first, because several gigabytes over your connection should always be your decision. --consent answers in advance, for a script.

• Weights come from the publisher, over HTTPS, from a URL pinned to an immutable revision. Nodeau never republishes weights, and you need no account or token.
• An interrupted download resumes.
• The SHA-256 is checked before the file is allowed to be used.
• A file that fails the check is moved aside rather than deleted, so the evidence survives and the next run can't mistake it for a good one.

nodeau run <model> offers to download a missing model for you, and --pull answers yes in advance.

The model cache#

--cache-dir points Nodeau somewhere else. It has to be the directory the runtime mounts.

Each machine keeps its own copy of a model and verifies it itself. A second machine downloads, or imports, what it needs, and Nodeau places work where the verified weights already are. Copying models between machines for you is on the roadmap.

Verification#

Presence isn't integrity, so the two are separate columns.

nodeau model verify qwen3.5-4b-q4km   # re-read this machine's copy and recompute
nodeau model status qwen3.5-4b-q4km   # every machine's view

On each machine
  NODE       CACHE     DETAIL
  nodeau-c   verified  -
  nodeforge  absent    this node has been observed and does not hold this model

Will it fit? (judged as if the card were free of Nodeau workloads)
  NODE       GPU                         VERDICT          NEEDS      AVAILABLE   HEADROOM
  nodeau-c   NVIDIA GeForce RTX 3080     yes (estimated)  4,406 MiB  9,365 MiB   4,959 MiB
  nodeforge  NVIDIA GeForce RTX 5060 Ti  yes              3,512 MiB  15,315 MiB  11,803 MiB

Only a computed SHA-256 that matches counts as verified, and only a verified copy is ever used to place a workload. Once, a file with the right name and the right size served fluent answers for a day while being the wrong bytes, which is why "it loads and sounds fine" doesn't count as evidence.

Each machine re-checks its own copies on its own: whenever a file's fingerprint changes, and again when a check gets old. nodeau model verify is the on-demand version, and it's worth running after a power cut or any storage trouble.

The fit columns answer is this card big enough, judged as if the card were free of Nodeau workloads. Whether it's free right now is shown separately, because "too small" and "busy" call for completely different next steps.

Removing a model#

nodeau model remove qwen3.5-4b-q4km

This deletes the weights on this machine only. Nodeau first checks that no service uses the model, and names the service if one does. If it can't reach the cluster to make that check, it stops rather than guessing, because not knowing whether something is in use is a reason to stop.

Other machines keep their copies. Deleting gigabytes on a machine you aren't sitting at is a decision you make there.

Measured and estimated#

Where a configuration has been measured on hardware like yours, Nodeau uses the measurement. Where it hasn't, it computes a figure from the model's own architecture and file size, adds a further margin, and labels the decision estimated. You always know which one you got.

nodeau model info <model>    # the measurements, and their scope

Each measurement row says whether it counted the whole device or one process, because those are different quantities, and mixing them would quietly weaken the fit check.

If a model is turned away only because of the extra margin for unmeasured hardware, you can accept that risk yourself with --accept-estimate-risk. See accepting the risk yourself.

Your own models#

Everything above is the curated set. To run a GGUF file of your own, see bring your own model. Your own models are listed separately from curated ones, because they're a different kind of trust: you chose them, their licence is yours to check, and you supplied the bytes.

These docs describe the current release, the beta channel. Spotted something wrong or missing? Tell us. A report from a machine we have never seen is the most useful thing we get.