Alright, so I am using Ubuntu and going through the terminal, I am trying to download a model, sometimes it works, but only gets to about 50% and then drops out.
Mostly I am getting an error, informing me that ‘pull model manifest: 429’ and the something about how I need an account or a hftoken.
Anyone got an idea on how to get around this?
If a 429 error is involved, there is no complete workaround. Mitigation may be possible using hf login or the HF_TOKEN environment variable.
Your error is Hugging Face rate limiting. Your “50% then drops” is the same root cause plus how the pull is being done on Linux.
When you run ollama pull … or ollama run …, Ollama first downloads a small metadata file called a manifest. The manifest tells Ollama which blobs or layers to fetch next.
If the manifest fetch fails, the whole pull stops early.
429 = Too Many Requests. The server you are pulling from is refusing more requests because you exceeded its rate limit in the last few minutes. Hugging Face documents this exactly: when you hit a rate limit you get HTTP 429, and the limit is tracked in a 5-minute window. (Hugging Face)
That message is the Hugging Face Hub telling you: “anonymous or unauthenticated IP traffic gets lower limits.” A token upgrades you from “anonymous per IP” limits to “logged-in account” limits, which are higher. Hugging Face even publishes the tier table (Anonymous vs Free vs Pro, etc.). (Hugging Face)
This exact Ollama symptom is reported by others too. Example: an Ollama issue where pull model manifest: 429 appears alongside Hugging Face token wording and an unexpected IP, which is consistent with shared egress IPs, VPNs, or carrier NAT. (GitHub)
If your command looks like this:
ollama run hf.co/<user>/<repo>
# or
ollama pull hf.co/<user>/<repo>
…then you are pulling from Hugging Face Hub. Hugging Face’s Ollama page explicitly documents this hf.co/{username}/{repository} format. (Hugging Face)
If you are pulling from the Ollama model library (no hf.co/...), you can still see 429 sometimes, but your “HF token” wording strongly suggests you are on the Hugging Face path.
Create a Hugging Face account and generate a User Access Token. Tokens can be scoped (read vs write). For downloads, a read token is the safer default. (Hugging Face)
This is the Ubuntu gotcha: Ollama often runs as a systemd service, so setting HF_TOKEN in your terminal may do nothing.
Ollama’s FAQ says: if Ollama runs as systemd, set env vars via systemctl edit ollama.service, then daemon-reload and restart. (Ollama Documentation)
Do this:
sudo systemctl edit ollama.service
Add:
[Service]
Environment="HF_TOKEN=hf_your_token_here"
Then:
sudo systemctl daemon-reload
sudo systemctl restart ollama
Why this works: Hugging Face documents HF_TOKEN as the environment variable that authenticates to the Hub and overrides any locally stored token. (Hugging Face)
Hugging Face rate limits are computed over the last 5 minutes. If you spam retries, you stay rate-limited longer. The rate-limit page also points you to a billing dashboard that shows your current bucket usage. (Hugging Face)
On locked-down networks, you can see mid-download failures and also weird manifest errors.
Ollama’s FAQ says:
HTTPS_PROXY for outbound model pulls.HTTP_PROXY because model pulls are HTTPS-only and HTTP_PROXY can interrupt client connections. (Ollama Documentation)If you need a proxy, set it the same systemd way:
sudo systemctl edit ollama.service
[Service]
Environment="HTTPS_PROXY=https://proxy.example.com:port"
# Do NOT set HTTP_PROXY for Ollama pulls
Then restart:
sudo systemctl daemon-reload
sudo systemctl restart ollama
Extra context: there are real reports that proxy env vars were not applied during manifest requests in some setups, producing manifest failures. (GitHub)
Download the GGUF with Hugging Face tooling, then import locally into Ollama.
Why this helps:
hf auth login) and its download tooling. (Hugging Face)hf auth login
Hugging Face’s CLI guide documents this as the normal login step. (Hugging Face)
For example, Hugging Face documents hf download <repo> as a standard way to download model content. (Hugging Face)
Ollama’s official import docs say: create a Modelfile with FROM /path/to/file.gguf, then ollama create …. (Ollama Documentation)
Example:
mkdir -p ~/ollama-import/mygguf
cd ~/ollama-import/mygguf
printf 'FROM /absolute/path/to/model.gguf\n' > Modelfile
ollama create my-model -f Modelfile
ollama run my-model
This is the cleanest approach if your network is flaky or you keep hitting 429 mid-pull.
Common reasons that fit your exact story:
HTTP_PROXY. (Ollama Documentation)systemctl status ollama --no-pager
If it is active as a service, use the systemd env var method above. (Ollama Documentation)
journalctl -u ollama -f
If you see 429 responses repeatedly, do token + slow retries.
HF_TOKEN env var definition: https://huggingface.co/docs/huggingface_hub/en/package_reference/environment_variables (Hugging Face)hf.co/... syntax: https://huggingface.co/docs/hub/en/ollama (Hugging Face)HTTPS_PROXY, avoid HTTP_PROXY): https://docs.ollama.com/faq (Ollama Documentation)FROM /path/to/file.gguf): https://docs.ollama.com/import (Ollama Documentation)pull model manifest: 429 is Hugging Face rate limiting in a 5-minute window. Token raises limits. (Hugging Face)HF_TOKEN in the systemd service, not only in your shell. (Ollama Documentation)HTTPS_PROXY and avoid HTTP_PROXY. (Ollama Documentation)hf auth login + hf download, then import GGUF locally via Modelfile. (Hugging Face)That was all no help, also why on earth go off about proxy’s and VPNs?
I have a network cable plugged from my modem directly into my server.
I have got this since yesterday without reasons. I really don’t undersand why HF do this. Reallys.
So.. What is the problem ?
It seems i have been banned just like that. Without nothing. (and as i don-t like that kind of injustice..)
You are trying to download the model from ollama or hugging face? If you are pulling the model from ollama, then it wouldn’t ask you for an HF token. If you’re trying to download from ollama you might just have to wait or create an ollama account and sign in through the terminal.
If you are downloading from hugging face then yeah, you get rate limited pretty heavily when trying to download without an account/auth - if you don’t have an HF token then I’d strongly suggest you get one. They are free and I’ve had zero issues downloading huge models.
As i say, i tried to pull model from HF. I have got an account with a HF token key (like you said) but, i am still block by 429 error. Even if i notice in my docker-compose the HF Token in the .env
I have got no problems pulling from Ollama (They do not block strangely, so..)
I have an IP which is rate limited despite i don’t try spamming. (One time yesterday, another this morning)
so the 5 minutes rules.. Euh..
If i do not had HF account, i could understand it. But that’s not the case.
huh - yeah that’s weird then. I’ve only seen a 429 error when downloading from an unauthenticated session. The other thing I’d suggest if you haven’t already is just generating a new HF token and trying with that one. Otherwise, reaching out directly to HF is probably your best move.
It seems to be good this morning… I don’t know what happened.. Strange case. ahah.
If i’ve got that thing again, i will reach HF directly.