Ollama + Kaggle + Cloudflare

Useful links

Before you start

• Run these sections in order, in the same Kaggle notebook session.
• Keep the notebook session alive. If Kaggle restarts, Ollama and the tunnel may stop, and the tunnel URL may change.
• The 27B Q4_K_M model is large. Downloading and loading it can take time and may exceed the runtime’s practical memory limits.
• A Quick Tunnel URL is temporary. Don’t treat it as a permanent production endpoint.

1. Install the required software

Use this once at the beginning of a fresh Kaggle runtime. It installs Ollama and Cloudflare Tunnel. The zstd package is a utility and may not be needed in every runtime.

!apt-get update -qq
!apt-get install -y zstd
!curl -fsSL https://ollama.com/install.sh | sh
!wget -q https://github.com/cloudflare/cloudflared/releases/latest/download/cloudflared-linux-amd64.deb
!dpkg -i cloudflared-linux-amd64.deb
!/usr/local/bin/ollama --version
!cloudflared --version

What to check: both version commands should print version information. If package installation fails, read the cell output before continuing.

2. Start the Ollama API server

Starts Ollama in the background and writes its logs to a file instead of suppressing errors. Run this cell once per runtime.

import subprocess
import time
import requests

ollama_server = subprocess.Popen(
    ["/usr/local/bin/ollama", "serve"],
    stdout=open("/tmp/ollama.log", "w"),
    stderr=subprocess.STDOUT
)

time.sleep(5)

response = requests.get("http://127.0.0.1:11434/api/tags", timeout=10)
print("Local Ollama status:", response.status_code)
print(response.text)

3. Download the model

Use ollama pull to download the model without opening an interactive chat session. This can take a long time because the model is large. Run it once per runtime unless the model is already stored.

! /usr/local/bin/ollama pull hf.co/theLittleStone/Qwen3.6-27B-AEON-Ultimate-Uncensored-MTP-GGUF:Q4_K_M
! /usr/local/bin/ollama list

If the download appears stuck, check the notebook output and disk and memory availability. Avoid starting an interactive ollama run session if you need the notebook to continue to later cells.

4. Start a fresh Cloudflare Quick Tunnel

Creates a temporary public HTTPS URL that forwards requests to the local Ollama API. This cell reads the tunnel log and extracts the generated URL automatically.

import re
import subprocess
import time

with open("/tmp/cloudflared.log", "w") as tunnel_log:
    cloudflared = subprocess.Popen(
        [
            "cloudflared", "tunnel",
            "--url", "http://127.0.0.1:11434",
            "--http-host-header", "localhost:11434",
            "--protocol", "http2"
        ],
        stdout=tunnel_log,
        stderr=subprocess.STDOUT
    )

tunnel_url = None
for _ in range(30):
    time.sleep(1)
    with open("/tmp/cloudflared.log", "r") as f:
        log_text = f.read()
    match = re.search(r"https://[a-z0-9-]+\.trycloudflare\.com", log_text)
    if match:
        tunnel_url = match.group(0)
        break

print("Tunnel process status:", cloudflared.poll())
print(log_text)
print("Tunnel URL:", tunnel_url)

Copy the newly printed URL if you need to enter it manually. Don’t reuse an old Quick Tunnel URL after a runtime or tunnel restart.

5. Test both API endpoints

Checks the local API and the public tunnel. The tunnel must be running for the public tests.

import requests

local_response = requests.get(
    "http://127.0.0.1:11434/api/tags", timeout=15
)
print("Local /api/tags:", local_response.status_code, local_response.text[:1000])

if tunnel_url:
    public_tags = requests.get(f"{tunnel_url}/api/tags", timeout=20)
    print("Public /api/tags:", public_tags.status_code, public_tags.text[:1000])

    public_models = requests.get(f"{tunnel_url}/v1/models", timeout=20)
    print("Public /v1/models:", public_models.status_code, public_models.text[:2000])
else:
    print("No tunnel URL was detected. Check /tmp/cloudflared.log.")

Expected result: HTTP 200 for /api/tags and, if supported by the installed Ollama version, /v1/models. Use the exact model ID returned by /v1/models in your client.

6. Connect Hermes or another compatible client

Use the current tunnel URL, not an old one.

Quick troubleshooting checklist

• Local API isn’t responding: check the Ollama process and read /tmp/ollama.log.
• Model pull fails or runs out of resources: confirm the model identifier, available disk space, and runtime memory.
• Tunnel URL is missing: read /tmp/cloudflared.log and verify that cloudflared is still running.
• Public API test fails: confirm the local endpoint works first, then confirm the tunnel process is alive and use the newest URL.
• Client can’t find the model: copy the exact model ID from /v1/models; don’t assume it matches the Hugging Face repository name.