Find and Pull a Model

tt knows which models run on your hardware, what each one needs, and where its weights come from. This lesson is about choosing; the next one serves what you chose.

What Runs Here

tt model list

tt model list

tt detects your board (n150 on the cloud pod, p300x2 on a TT-QuietBox 2) and lists the models that have a serving profile for it, with the engine and where the weights are. tt model list --catalog shows only the released catalog, the models Tenstorrent ships and tests through tt-inference-server. --type llm|image|audio|video narrows by kind, and --all shows every model on every device.

tt model list --catalog

tt model list --catalog

What a Model Needs

tt model info Llama-3.1-8B-Instruct

tt model info Llama-3.1-8B-Instruct

│ hf_repo       │ meta-llama/Llama-3.1-8B-Instruct                             │
│ type          │ llm                                                          │
│ engines       │ vLLM, forge                                                  │
│ via           │ inference-server, studio                                     │
│ servable      │ yes — `tt serve Llama-3.1-8B-Instruct` (`--studio` deploys   │
│               │ it with TT-Studio)                                           │
│ parameters    │ 8B                                                           │
│ min disk      │ 36 GB                                                        │
│ min ram       │ 20 GB                                                        │
│ cached        │ yes (15.0 GB)                                                │
│ on n150       │ vLLM — COMPLETE, max context 65536                           │
│ on n300       │ vLLM — COMPLETE, max context 131072                          │
│ on t3k        │ vLLM — COMPLETE, max context 131072                          │
│ ...           │                                                              │

cached says the weights are already in the Hugging Face cache on this machine (captured on a TT-QuietBox 2 that had pulled them earlier).

Validation status goes EXPERIMENTAL, FUNCTIONAL, COMPLETE. COMPLETE is fully validated; FUNCTIONAL runs correctly and is not fully tuned. The status is per board, so read the row for yours. min disk and min ram are in the same table.

Starting points

Gated models (Llama, Gemma) need you to accept the license on the model's Hugging Face page first. The hf_repo line in tt model info names the page.

Log In to Hugging Face

Serving needs a Hugging Face token, even for models that are free to download, because tt-inference-server checks it before it starts. Log in once:

tt model login

tt model login

Choose Log in with your browser (it shows a URL and a short code) or paste a read token from huggingface.co/settings/tokens. tt finds the token again on every later run, and tt serve --dry-run tells you where it came from without printing it.

Pull Weights Ahead of Time

tt serve fetches weights on first use. To download now, while you read on:

tt model pull Llama-3.1-8B-Instruct

tt model pull Llama-3.1-8B-Instruct

The weights land in the Hugging Face cache and tt reports where. On the cloud pod that cache is ~/models (HF_HOME), which hf, the vLLM plugin and tt all share.

Community Models

The catalog is what Tenstorrent ships and tests. Beyond it, anyone can publish a model bundle to the Hugging Face Hub with tt-model-manager: a bring-up packaged so that someone else can serve it with the same commands. A bundle carries the engine the author built and a pointer to the weights, which download into your own cache under your own token.

⚠️ Community models are not reviewed by Tenstorrent. Serve at your own risk. Bundles are published by their owners. Tenstorrent does not test or maintain them, and a bundle can run code and download weights you have not looked at. Read tt model info and the author's page before you pull one, and do not point one at data you would not trust to a stranger's software. The compatibility check only confirms the chip architecture; it says nothing about whether the model is safe.

To browse instead of search, the tt-awesome model browser lists every catalog model and community bundle by collection, family, kind and hardware, with the commands to install and run each. It marks community bundles with the same warning.

tt model search qwen3 --limit 5     # find published bundles on the Hub
tt model list --community           # community bundles that run on this machine
tt model info raahemnabeel/qwen3-coder-next    # manifest and compatibility verdict

tt model search qwen3 --limit 5

A real search on a TT-QuietBox 2 (2026-10-28):

          tt-model bundles on the Hub — every published bundle matching 'qwen3'
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━┓
┃ name                                ┃ visibility ┃ downloads ┃ updated    ┃ installed ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━┩
│ raahemnabeel/qwen3-5-122b-a10b      │ public     │ 0         │ 2026-10-28 │ —         │
│ raahemnabeel/qwen3-coder-next       │ public     │ 0         │ 2026-10-28 │ ✓         │
│ episod/qwen3.8-27b-dflash2-p300     │ public     │ 0         │ 2026-10-28 │ ✓         │
└─────────────────────────────────────┴────────────┴───────────┴────────────┴───────────┘

Names are namespace/name. A name with a slash that is not a catalog model is a bundle, and tt serve routes it to tt-model-manager. Compatibility is checked, not guaranteed: pulling or serving a bundle checks that its binaries were built for your chip architecture and warns if your box has fewer chips than the profile needs.

"I want to..." and what to search

These are all community bundles, so the warning above applies to every row. tt model search matches the text of a bundle's name, so search for a model family or a task word that models usually carry in their name. These all returned bundles on 2026-10-28; bundles are published often, so your results will differ.

Some searches found nothing (whisper, video, robot, pose, rerank). Speech-to-text and video models such as whisper-large-v3 and Wan2.2-T2V-A14B-Diffusers are catalog models, so look for them with tt model list --type audio or --type video. The other terms are just not in any bundle's name yet. Empty results mean nobody has published one.

The suffix on a bundle name is the hardware it was built for: -n150 or -p150 is one chip, -p300 is one board of two chips, -p300x2 is a whole TT-QuietBox 2. tt model list --community --hw <board> shows only the bundles that fit your machine.

Downloading by Hand

hf download is for a repository you will use outside tt, such as the TTNN demos in Build on the Stack. The cloud pod has hf installed; elsewhere, uv tool install huggingface_hub.

mkdir -p ~/models
hf download Qwen/Qwen3-0.6B --local-dir ~/models/Qwen3-0.6B          # 1.5 GB, ungated
hf download meta-llama/Llama-3.1-8B-Instruct --local-dir ~/models/Llama-3.1-8B-Instruct   # 16 GB, gated

Check space first: df -h ~/models. A partial download leaves the directory incomplete; re-run the same command to resume.

Next: Serve a Model →