[Presentation Materials] I Gave a Talk at Midosuji Tech Titled "I Tried Running Decision Model's Cloudflare 'Clef-Flash' Locally" #midosuji_tech

Hello, I'm Arahira (@eiraces) from the Professional Services Division, Cloud Business Headquarters.

I gave a presentation titled "I Ran Cloudflare's Decision Model 'Clef-Flash' Locally" at "Classmethod Osaka IT Study Session Midosuji Tech ~Trying Out Decision Models Edition~" held on Thursday, October 29, 2026, so I'm publishing the slides.

https://classmethod.connpass.com/event/409980/

Presentation Slides

Additional Notes

Since the theme this time was "Lightning Talks on Trying Out Decision Models," I decided to run Clef-flash, announced by Cloudflare, on my own Mac.
It's published on Hugging Face, so you can download it directly with Ollama and use it.

The test environment was a MacBook Pro (Apple M4 Pro, 48GB unified memory), with Ollama 0.40.11.
The nominal VRAM requirement for Clef-flash is over 41GB, so I was worried about whether it would barely run, but by using a quantized model (about 12GB) it ran smoothly without any problem.

When I first asked "are you ai?", the response was "yes 95%".
A Decision model doesn't generate sentences but returns only a typed answer and a probability, so this terseness is normal behavior.
By the way, to "are you a flower?" it answered confidently with "no 99%". No matter how many times I tried, it returned roughly the same values.

In this LT, I prepared and presented two demos.

• Receive prompts with Claude Code's UserPromptSubmit hook, classify them into implementation requests, research requests, Q&A, etc., and aggregate them
• Periodically obtain the active window's app name and title, classify them into focus, chores, and breaks, and aggregate the day's time

The reason I chose Clef-flash is that I wanted both to be completed locally.
Prompts and window titles in particular are a mass of confidential information, so it's reassuring not to have to send them to an external API.

Through the testing for this LT, my impression was that it isn't suited to use cases that require reliable classification.
The confidence for the "chores" classification dropped to around 0.1 during the time I had Google Calendar or Slack open, so I made it show up in the report as a case where judgments tend to waver.
Also, when there's a tie in Choice, the first option is selected, so you need to pay attention to the order of the options as well.

It seems sufficiently usable for data preprocessing or situations where you want to attach rough labels, so I plan to keep it running resident for a while and see how it goes.
The Q&A session on the day was lively, with many interested people, which was great.

I hope this entry helps someone.
This has been Arahira from the Professional Services Division, Cloud Business Headquarters!

References

https://huggingface.co/Cloudflare/clef-flash