Maintainer of this skill?
Automated security scan — not a certification.
by ARahim3
Frontier-class open models on a free Kaggle TPU v5e-8: GLM-5.3-Flash 320B MoE (~165 tok/s, our own JAX engine) and Qwen3.8-27B bf16 (~130 tok/s), 262k context, prefix caching. Works with Claude Code, Codex, opencode and pi.
# Add to your Claude Code skills
git clone https://github.com/ARahim3/kaggle-tpu-labGuides for using ai agents skills like kaggle-tpu-lab.
Last scanned: 9/5/2026
{
"issues": [],
"status": "PASSED",
"scannedAt": "2026-09-05T08:05:20.325Z",
"npmAuditRan": true,
"pipAuditRan": true,
"promptInjectionRan": true
}See how kaggle-tpu-lab compares with popular alternatives.
kaggle-tpu-lab is an open-source ai agents skill for AI coding assistants such as Claude Code, Codex CLI, and ChatGPT, built by ARahim3. Frontier-class open models on a free Kaggle TPU v5e-8: GLM-5.3-Flash 320B MoE (~165 tok/s, our own JAX engine) and Qwen3.8-27B bf16 (~130 tok/s), 262k context, prefix caching. Works with Claude Code, Codex, opencode and pi. It has 688 GitHub stars.
Yes. kaggle-tpu-lab passed SkillsLLM's automated security scan — a dependency vulnerability audit plus prompt-injection heuristics — with no high-severity issues. You can read the full report in the Security Report section on this page.
Clone the repository with "git clone https://github.com/ARahim3/kaggle-tpu-lab" and add it to your Claude Code skills directory (see the Installation section above).
kaggle-tpu-lab is primarily written in Python. It is open-source under ARahim3 on GitHub, so you can review or fork the full source.
Yes. SkillsLLM lists many other AI Agents skills you can browse and compare side by side. Open the AI Agents category from the badge at the top of this page, or use the Related Skills and comparison links further down to weigh kaggle-tpu-lab against similar tools.
No comments yet. Be the first to share your thoughts!
30 days in the Featured rail · terms & refunds
⚠️ Third-Party Software Notice
This skill is third-party open-source software developed and hosted independently on GitHub. SkillsLLM is an informational directory and does not control or maintain the underlying repository.
Any security checks, ratings, or warnings displayed by SkillsLLM are automated and limited in scope. They do not constitute a security certification or guarantee that the software is safe, error-free, or free from malicious code, vulnerabilities, compromised dependencies, or prompt-injection risks.
Review the source code, permissions, dependencies, and configuration before installing or running any third-party skill. Use is at your own risk. To the maximum extent permitted by applicable law, SkillsLLM is not liable for losses arising from third-party software.
Automated security scan — not a certification.
Run open models on Kaggle's free TPU v5e-8 and get a public endpoint that speaks the OpenAI and Anthropic APIs. Point Claude Code, Codex CLI, opencode or anything else at it. No GPU, no cloud bill, about twenty minutes from pressing Run to a URL.
Each model has its own folder with a run-all Kaggle notebook, the kernel script behind it, and a write-up of how it works and what we measured.
| Model | Weights on the TPU | Context | One stream | Many streams | Prefill | Run → URL | Engine | |
|---|---|---|---|---|---|---|---|---|
| Qwen3.8-27B | bf16, no quantization | 262k | ~130 tok/s | ~540 tok/s at 8 | 10,300 tok/s | ~22 min | vllm-tpu + one patch | notebook |
| GLM-5.3-Flash (320B MoE) | 3-bit experts, int8 rest | 262k | ~165 tok/s | ~170 tok/s at 3 | ~1,800 tok/s | 6–13 min | our own JAX engine | notebook |
Numbers are measured on the shipped configuration; the folder READMEs say how. Qwen runs on vllm-tpu with one patch. GLM-5.3-Flash runs on an engine we wrote in JAX for it; as far as we know it is the first to run that model on a TPU.
A Kaggle account with TPU access (verify it under Settings: a phone number, and for some accounts an identity check through Persona before TPU sessions are handed out) and its free quota, around 20 TPU hours a week. Nothing to install for the notebook route. For the terminal route, Python 3.9+ and the Kaggle CLI.
The notebook attaches public datasets holding the weights (and, where it helps, a pre-built compile cache), builds the engine across the eight chips, opens a Cloudflare tunnel and prints the URL and an API key. The last cell is the server: leave it running. A keepalive holds the session up to Kaggle's limit, a little under nine hours, after which you run it again and get a new URL.
Wire a coding agent with one line, for example Claude Code:
ANTHROPIC_BASE_URL=<url> ANTHROPIC_AUTH_TOKEN=<key> ANTHROPIC_MODEL=<model> claude
Each folder README has the exact lines for Claude Code, Codex CLI and opencode.
git clone https://github.com/ARahim3/kaggle-tpu-lab
cd kaggle-tpu-lab
python launch.py serve # Qwen3.8-27B today; --model picks another recipe once there is one
launch.py pushes the kernel with the Kaggle CLI and follows its progress; status
and stop do what they say.
The GLM-5.3-Flash recipe also runs on a rented TPU VM, the same machine as Kaggle's (v5litepod-8) or a v6e-8,
with no queue and no nine-hour cap: gcp/ has the guide and the three scripts (environment, datasets, run).
One folder at the top level, named after the model: README.md with the numbers and the
how, kernel/ with the serving script, notebook/ with the run-all notebook generated
from it, plus whatever the recipe needs (a patch, an engine). The launcher and the
notebook share the same config block, so a setting changed in one place means the same
thing in the other.
The code here is MIT. Model weights keep their own licenses; each folder says which.