# How to use Kaggle's free TPU v5e-8: a 128 GB AI machine, step by step

> Every click from creating an account to confirming the TPU works, each with a screenshot, then two ready-made ways to run Qwen 27B or Gemma 26B on it. Plus tips to make your quota last, and the Kaggle rules that can get an account banned.

By Technical Paji (@technicalpaji). Updated 1 Nov 2026. Facts checked against official docs on 1 Nov 2026.
Canonical: https://technicalpaji.pages.dev/tpu/

## Quick answer

Yes. Kaggle's free notebooks include a **TPU v5e-8** accelerator: **8 TPU chips with 16 GB of memory each, 128 GB in total**. Kaggle's docs say you can use TPUs for **up to 20 hours a week** and **up to 9 hours in a single session**. You need a phone-verified account, but no credit card.

## What you get

- **Accelerator:** TPU v5e-8: 8 TPU chips, each with 16 GB of HBM memory [2]
- **Total memory:** 8 × 16 GB = 128 GB. In one public run, each chip had 15.75 GiB usable, 125.98 GiB in total [6]
- **Quota:** Up to 20 hours of TPU per week [1]
- **Session:** Up to 9 hours in a single session [1]
- **Where:** Notebook menu: `Settings > Accelerator > TPU v5e-8` [3]
- **GPU option:** GPU T4 ×2 is also available. The GPU P100 was retired on 15 Sep 2026 [4]

## Before you start

- A Kaggle account. You can register with Google or with an email address.
- A phone number that can receive a text message, to verify the account. Without it, notebooks get no TPU and no internet.
- One account per person. Kaggle may ban duplicate accounts immediately.
- A browser and about 10 minutes. Everything runs on Kaggle; nothing is installed on your computer.

## Part 1: Set up your Kaggle account

### Step 1: Create a free Kaggle account

Click: `kaggle.com` → **Register**

Choose **Register with Google** or **Register with Email**, whichever is easier. No card is needed.

Screenshot: Kaggle's Welcome page with the Register tab open, showing the 'Register with Google' and 'Register with Email' buttons. (kaggle.com → **Register** → Google or email)

### Step 2: Verify your phone number

Click: **Profile picture** → **Settings** → **Phone verification**

Enter your number, then the code Kaggle texts you. When it's done, the section says **Verified**. Without this, notebooks get no TPU, no GPU and no internet.

Screenshot: The Phone verification section of Kaggle's settings, showing 'Verified'. (Settings → Phone verification → **Verified**)

> **One account per person.** Kaggle's community guidelines say **immediate account bans may be issued** to people who create duplicate accounts. Don't make a second account for more quota.

## Part 2: Turn on the TPU and check it

### Step 3: Create a new notebook

Click: **+ Create** → **Notebook**

Click the **+ Create** button on the left, then **Notebook**. A blank notebook opens.

Screenshot: Kaggle's '+ Create' menu, open, listing Notebook, Dataset, Competition and Benchmark. (**+ Create** → **Notebook**)

### Step 4: Choose TPU v5e-8 as the accelerator

Click: **Settings** → **Accelerator** → **TPU v5e-8**

Open **Settings** in the menu at the top of the notebook. Some screens call it **TPU VM v5e-8**; it's the same thing. The notebook restarts, and getting a TPU can take a few minutes.

**Internet** must be on too, to download models. In the same menu, click **Turn on internet** if you see it. If it says **Turn off internet**, internet is already on.

Screenshot: The Settings menu of a Kaggle notebook: Turn off internet, Accelerator and Environment Preferences. The Accelerator submenu lists None, GPU T4 ×2, GPU P100 and TPU v5e-8, with TPU v5e-8 ticked. (Settings → Accelerator → ✓ **TPU v5e-8**. This screenshot is older: the GPU P100 was retired on 15 Sep 2026 and no longer appears in the list. Screenshot: [Kitkitkittt/kaggle-tpu-lab](https://github.com/Kitkitkittt/kaggle-tpu-lab))

> **You may have to wait for a TPU.** Kaggle's docs say you may be placed in a queue for a TPU at busy times. Wait a little, or try again later. Only one TPU session runs at a time, so stop any other TPU notebook first.

### Step 5: Check that all 8 chips are there

Type this in the first cell and press **Shift + Enter**:

```python
import jax
print(jax.device_count())
print(jax.devices()[0].device_kind)
```

Output:

```text
8
TPU v5 lite
```

If you see **8** and **TPU v5 lite**, all 8 TPU chips are attached. JAX calls the TPU v5e "TPU v5 lite".

> **Got 1 instead of 8?.** If the output shows `1` and `cpu`, no TPU was attached. Stop the session, choose **TPU v5e-8** under **Accelerator** again, and rerun the cell. If `import jax` itself fails, run `!pip install -q "jax[tpu]"` first.

## Part 3: Option A: run Qwen 27B (ready-made notebook)

Now run a large AI model on the TPU. There are two ways; pick one. **A** is the easiest: a ready-made notebook that gives you your own API. In **B**, Gemma 26B answers right inside the notebook. Both are public projects that have run on this same free TPU. The speed figures come from their READMEs; your run may differ a little.

### Step 6: Copy the public Qwen notebook

Click: **Open the notebook** → **Copy & Edit**

Open Kiet Vo's public notebook: [Qwen3.8-27B-bf16 for free on Kaggle TPU with API](https://www.kaggle.com/code/keithvo/qwen3-8-27b-bf16-for-free-on-kaggle-tpu-with-api). Click **Copy & Edit** at the top; your own copy opens.

In the copy, choose **TPU v5e-8** under **Accelerator** and turn internet on, as in step 4.

### Step 7: Check that both datasets are attached

Click: **Right panel** → **Input**

Open **Input** in the notebook's right panel. These two datasets should be there. They come attached with the copy; if one is missing, add it with **Add Input** by searching for its name.

- `rahim3/qwen3-8-27b-bf16`: the model, 55.6 GB
- `rahim3/qwen38-tpu-env-v5e8`: a ready-made compile cache

### Step 8: Click Run All and wait about 22 minutes

Click: **Run All**

The project's README says loading and compiling the model takes about 22 minutes. The last cell is the server: when it's ready, it prints a banner with a **Base URL** (`…trycloudflare.com/v1`), an **API key** (`sk-…`) and the model name `qwen3.8-27b`.

### Step 9: Talk to the model

Put the Base URL and API key into any OpenAI-compatible app, or run this curl command in a terminal. Replace the URL and key with your own:

```bash
curl https://xxxx-yyyy.trycloudflare.com/v1/chat/completions \
  -H "Authorization: Bearer sk-<your-key>" \
  -H "Content-Type: application/json" \
  -d '{"model":"qwen3.8-27b","messages":[{"role":"user","content":"Hi! Explain a TPU in one line."}]}'
```

According to the README: about 130 tokens/s for one user, about 540 tokens/s in total for 8 users at once, and a context of up to 262K tokens.

> **This URL is for your own testing only.** Anyone with the URL and API key can use your TPU, so don't share them. Don't use it as the server for an app or website either: Kaggle's Acceptable Use Policy calls **server farming** an abuse of its resources, and it can get your account banned. Stop the session as soon as you're done.

## Part 4: Option B: run Gemma 26B (JAX)

Here Google's Gemma 4 26B model answers right inside the notebook. No public URL is created.

### Step 10: Create a Hugging Face read token

Click: `huggingface.co` → **Settings** → **Access Tokens** → **Create new token**

Create a free Hugging Face account, then create a token of type **Read** and copy it. The project's setup uses this token to download the model.

### Step 11: Add the token to Kaggle Secrets

Click: **Add-ons** → **Secrets** → **Add Secret**

Enter `HF_TOKEN` as the label, exactly as written, and your token as the value. After saving, tick the box next to the secret so this notebook can read it.

This notebook also needs **TPU v5e-8** and internet turned on, as in step 4.

### Step 12: Run the setup (first time only)

Run this in a new cell:

```bash
%%bash
cd /root && rm -rf kaggle-tpu-inference && \
git clone https://github.com/YLiu95/kaggle-tpu-inference.git && \
cd kaggle-tpu-inference && bash setup.sh
```

According to the README, `setup.sh` checks that JAX sees 8 chips, downloads the ~52 GB model and builds a compile cache (about 7 minutes, once).

### Step 13: Ask Gemma a question

Run this cell as often as you like; just change the question in quotes:

```bash
%%bash
cd /root/kaggle-tpu-inference && \
bash run.sh "Hi! Explain a TPU in one line." --max-new-tokens 768 --temperature 1.0
```

According to the README: the first token arrives in 44-69 ms, then about 111 tokens/s. The model is split across the 8 chips, using only about 6 GB on each.

Screenshot: Terminal dashboard 'TPU utilisation': chip 0, TPU v5 lite, 6.12 GiB used out of 15.75 GiB. Aggregate: HBM 48.96 / 125.98 GiB over 8 chips. (Along with the answer, `run.sh` shows this TPU dashboard. This is the run in the project's README: each chip is a **TPU v5 lite** with 15.75 GiB, **125.98 GiB over 8 chips** in total (8 × 16 GB = 128 GB, with a little reserved by the system). Source: [YLiu95/kaggle-tpu-inference](https://github.com/YLiu95/kaggle-tpu-inference))

## Make your free quota last

- **Done? Stop the session.** An open session uses quota even when no code is running. Closing the tab isn't enough: Kaggle's docs say an idle session can take up to 60 minutes to stop by itself. Stop it yourself.
- **A session lasts up to 9 hours.** For long jobs, keep saving output to `/kaggle/working`. Everything else is deleted when the session ends.
- **One TPU session at a time.** If a TPU notebook is running in another tab, you won't get a new TPU.
- **The first run is slow, later ones are fast.** The model compiles the first time. The Qwen notebook comes with a ready-made compile cache, and after setup the Gemma project answers every question in the same session straight away.

## TPU or GPU: which should you pick?

| Job | Pick |
|---|---|
| Running large LLMs such as Qwen and Gemma (vLLM-TPU, JAX), or training and fine-tuning your own model (JAX, PyTorch/XLA, TensorFlow) | **TPU v5e-8**: 128 GB, up to 20 hours a week |
| Tools built for NVIDIA GPUs (CUDA), such as Ollama, llama.cpp, ComfyUI and Stable Diffusion | **GPU T4 ×2**: 2 × 16 GB, 30 hours a week or more |

In short: for a big model that needs lots of memory, use the TPU. For ready-made GPU apps, use T4 ×2. Both are free, and each has its own quota.

## If something goes wrong

- **Only cpu:0 shows up, or "TPU topology check FAILED".** No TPU was attached. Stop the session right away (to save quota), choose **TPU v5e-8** yourself under **Session options**, and run again. If it still fails, check that your phone is verified, you have TPU quota left, and no other TPU session is running.
- **Can't select the TPU, or stuck in a queue.** There's a queue at busy times; try again a bit later. Also check that your phone is verified.
- **"Device or resource busy" after a crash.** The chips stay busy for 10-30 seconds after a process ends. Wait a moment, then run `jax.devices()` again.
- **"TPU not found" in TensorFlow.** TensorFlow needs a different setup on v5e-8. See this [public notebook](https://www.kaggle.com/code/herbison/tensorflow-v5e-8-oct-2025).
- **Model download fails.** Is internet on in Settings? For Gemma, is the `HF_TOKEN` secret named exactly that, and ticked for this notebook?
- **Don't use it as a server.** Kaggle's Acceptable Use Policy lists **server farming**, crypto-mining and activity unrelated to ML data science as abuse of its resources, and its community guidelines say abusing resources can lead to an immediate ban. Use the TPU to learn, train and test models.

## Frequently asked questions

### Is Kaggle's TPU really free?

Yes. Kaggle's TPU docs say TPUs are available for free. You need a phone-verified account, and use is limited by a weekly quota.

### How many hours do I get?

Kaggle's docs say up to 20 hours a week, and up to 9 hours in a single session.

### Is it one 128 GB chip?

No. TPU v5e-8 has 8 separate chips with 16 GB each. To run a big model you have to split it across the 8 chips, which both projects above do for you.

### Why is the TPU option greyed out or missing?

Usually your phone number isn't verified yet. Verify it in Settings and reload the notebook. At busy times you may also have to wait in a queue for a TPU.

### Will Ollama or Stable Diffusion run on it?

Those tools are built for NVIDIA GPUs (CUDA) and don't run on a TPU. For them, choose GPU T4 ×2 under Accelerator.

### Is the GPU P100 still available?

No. Kaggle retired the P100 on 15 September 2026. T4 ×2 is the only GPU option now.

### Can I run my app or website on it?

No. Kaggle's Acceptable Use Policy calls server farming an abuse of its resources, and sessions end after at most 9 hours anyway.

### Why do I have to download everything again each session?

Kaggle only saves what's in /kaggle/working (up to 20 GB). The rest of the disk is wiped when the session ends.

## Sources

1. [Tensor Processing Units (TPUs) Documentation](https://www.kaggle.com/docs/tpu), Kaggle documentation
2. [TPU v5e](https://docs.cloud.google.com/tpu/docs/v5e), Google Cloud documentation
3. [New TPU v5e-8s! Phasing out v3-8s](https://www.kaggle.com/discussions/product-announcements/607202), Kaggle staff announcement
4. [Sunsetting the NVIDIA Tesla P100 GPU on September 15, 2026](https://www.kaggle.com/discussions/product-announcements/735239), Kaggle staff announcement
5. [Notebooks](https://www.kaggle.com/docs/notebooks), Kaggle documentation
6. [Gemma-4 26B-A4B on a Kaggle TPU v5e-8 (kaggle-tpu-inference)](https://github.com/YLiu95/kaggle-tpu-inference), YLiu95 on GitHub
7. [Qwen3.8-27B on Kaggle's free TPU v5e-8 (kaggle-tpu-lab, from ARahim3's original)](https://github.com/Kitkitkittt/kaggle-tpu-lab), Kitkitkittt on GitHub
8. [Qwen3.8-27B-bf16 for free on Kaggle TPU with API](https://www.kaggle.com/code/keithvo/qwen3-8-27b-bf16-for-free-on-kaggle-tpu-with-api), Kaggle notebook by Kiet Vo
9. [google/gemma-4-26B-A4B-it](https://huggingface.co/google/gemma-4-26B-A4B-it), Hugging Face
10. [Efficient GPU Usage Tips](https://www.kaggle.com/docs/efficient-gpu-usage), Kaggle documentation
11. [Acceptable Use Policy](https://www.kaggle.com/aup), Kaggle
12. [Kaggle Community Guidelines](https://www.kaggle.com/community-guidelines), Kaggle
13. [tensorflow v5e-8 oct 2025](https://www.kaggle.com/code/herbison/tensorflow-v5e-8-oct-2025), Kaggle notebook by herbison