Bertaux Florian

Passionate about emerging technologies, cybersecurity (covering Red, Blue, and Grey teams), Artificial Intelligence, and software development.

[Bertaux Florian @_LEFBE]

Ka1zen — chat with open LLMs 100% offline on your Mac.
🧠 Qwen, Gemma, DeepSeek, Mistral, Llama…
🖼 Vision · 🎙 Audio · 🌐 Web search
🎨 FLUX image generation · 📚 RAG
🔒 Zero cloud, zero telemetry

Powered by Apple MLX.

https://t.co/7lOijJotiB

[Bertaux Florian @_LEFBE]

Ka1zen generation stats
Model: mlx-community/Qwen3.6-35B-A3B-4bit
Tokens: 2753 → 4375 (input → output)
TTFT: 831 ms
Speed: 163.4 t/s generation

Not bad :)

[Bertaux Florian @_LEFBE]

Excellent, big valid update, now I'm switching to full MTP support associated with real use cases, not marketing benchmarks (Websearch, tools, offline prompt, image generation, code)...

[Bertaux Florian @_LEFBE]

Enhancement in progress:
Ka1zen generation stats
Model: bartowski/Qwen_Qwen3.6-35B-A3B-GGUF
Tokens: 6658 → 1566 (input → output)
TTFT: 3.02 s
Speed: 104.1 t/s generation

[Bertaux Florian @_LEFBE]

Ka1zen generation stats (Websearch)
Model: bartowski/Qwen_Qwen3.6-35B-A3B-GGUF
Tokens: 7966 → 1575 (input → output)
TTFT: 3.68 s
Speed: 90.9 t/s generation

Model: ggml-org/gemma-4-12B-it-GGUF
Tokens: 850 (output)
TTFT: 4.86 s
Speed: 51.1 t/s generation

[Bertaux Florian @_LEFBE]

I'm currently working on a major update for Ka1zen.

Stay tuned!

[Bertaux Florian @_LEFBE]

Let's go! Let's verify if gemma-4-12B-it-bf16 works well on Ka1zen
Then I will test with MTP once available.

[Bertaux Florian @_LEFBE]

@ivanfioravanti My pleasure :)

[Bertaux Florian @_LEFBE]

@ivanfioravanti Qwen 27B

[Bertaux Florian @_LEFBE]

@ivanfioravanti This type of prompt allowed my son to better understand the solar system in a fun way. (qwen3.6:35b-a3b-coding-mxfp8)

[Bertaux Florian @_LEFBE]

@bridgemindai You can't compare a single machine (even a powerful one) to a high-performance server farm.
In my view, local models are designed to serve individuals, not businesses; otherwise, you'd need servers or more machines.

[Bertaux Florian @_LEFBE]

Benchmark scores often impress, but they rarely mention quantization. Mine use Q8, not Q4. Q4 may perform better, but it’s not production-ready

[Bertaux Florian @_LEFBE]

@karpathy What a great move! I'm jealous :)
I wish you all the best in such excellent company

[Bertaux Florian @_LEFBE]

Ka1zen v0.3.43 — both your asks landed :
– Help → Check now opens the banner (no more popup duplication)
– Auto-installer cleans up the DMG after relaunch

Plus : Favourites section in the model picker + colorised Update prerequisites log with collapsed pip noise

[Bertaux Florian @_LEFBE]

@jun_song I admit that keeping up with the integration of new models and tests into Ka1zen isn't easy.

[Bertaux Florian @_LEFBE]

Ka1zen now ships the full Qwen-Image bundle 🎨

– Qwen-Image T2I + Qwen-Image-Edit
– Edit-prompt optim : EN translate + negative-flip + identity cues
– Per-image provenance badge
– Plus : Server Console + reasoning-model stability fixes
https://t.co/f1kE6EMG8Y
@Alibaba_Qwen

[Bertaux Florian @_LEFBE]

It's frustrating, even with my machine, there are models where I can't test the implementation. 😭

[Bertaux Florian @_LEFBE]

When mlx-lm is updated by @ml_explorer , I will test the implementation, in Ka1zen, of the following models:
- Ling-2.6-flash-mlx-4bit-DWQ
- DeepSeek-V4-Flash-2bit-DQ

[Bertaux Florian @_LEFBE]

I'm working on implementing Qwen-Image-Edit!

This will allow Ka1zen to offer two complete image generation and editing solutions.

- Flux2.Klein
- Qwen Family - Qwen Image and Qwen Image Edit

[Bertaux Florian @_LEFBE]

@jun_song My testing protocol with Ka1zen is simple: never trust benchmarks. Always test under real-world conditions, on real-world use cases. If it works correctly, then I implement it; otherwise, I forget about it.