Laurent Zuijdwijk

Ai research at Mountains + code

[Laurent Zuijdwijk @laurent_zw]

@ItsmeAjayKV just prompt cerebras at 1000t/s or openrouter?

[Laurent Zuijdwijk @laurent_zw]

@jauneclaude Exactly this. My strix halo has been at 100% GPU power for days now and hardly uses any power. I can still work on it just fine

[Laurent Zuijdwijk @laurent_zw]

@yashmp2004 There is plenty of that around

[Laurent Zuijdwijk @laurent_zw]

@the_jimmy_jones @OpenAI I do that all the time haha

[Laurent Zuijdwijk @laurent_zw]

Not sure I like the naming for DeepSeek V4.1.

It seems like it is a new architecture and much bigger than V4.

Calling it V5 would've worked better

[Laurent Zuijdwijk @laurent_zw]

@jauneclaude @0xSero

[Laurent Zuijdwijk @laurent_zw]

@jauneclaude @0xSero bit big? 😅

[Laurent Zuijdwijk @laurent_zw]

@Italianclownz Thanks a lot! It is just too much fun to create all this stuff

[Laurent Zuijdwijk @laurent_zw]

Signal 3.8 27B just drew one of the best pelican-on-a-bike SVGs I've seen (Q4_XL quant). The fish on the luggage rack is a new one to me 😂

Qwen 3.5 27B took ~10% more tokens for a blander result.

Feels like tighter, more focused reasoning beats meandering on output and it's faster too.

Model used: https://t.co/bQOyFDjsM8

[Laurent Zuijdwijk @laurent_zw]

@teale_ai I am running more tests, but it is looking good so far

[Laurent Zuijdwijk @laurent_zw]

Qwen3.8 27B — the local inference favourite just got better and faster. 🚀

Launching Signal 3.8 27B, a Qwen3.8 autodistillation.

⚡ Up to 57% fewer tokens
🕒 Under half the wall clock time
✅ Matches or improves accuracy
📦 Drop-in AP quants for llama.cpp

https://t.co/bQOyFDjsM8

[Laurent Zuijdwijk @laurent_zw]

@notrajivpoddar Yep, it helps a lot, making thinking more terse

[Laurent Zuijdwijk @laurent_zw]

@woldpt I had one Q3 quant but it was bigger than the Q4XS. If I have time I will add a smaller quant. What is your target size?

[Laurent Zuijdwijk @laurent_zw]

@0xBakeer That matches my experience with my finetune of qwen3. 8 27b. Making it more terse makes Mtp faster.

[Laurent Zuijdwijk @laurent_zw]

@TeksEdge That doesn't sound right

[Laurent Zuijdwijk @laurent_zw]

@BradleyBenner Cool. If you spot any weirdness or loss of fidelity please let me know. ❤️

[Laurent Zuijdwijk @laurent_zw]

Not statistically significant, but signal scores better than the base model. It thinks more effectively

[Laurent Zuijdwijk @laurent_zw]

@boxwrenchx If you get to try it out, it would be great to hear how it works for you

[Laurent Zuijdwijk @laurent_zw]

I would love to hear community feedback. If there are any real life issues I would like to know, there are some adjustments I could make if needed. 🧡

[Laurent Zuijdwijk @laurent_zw]

Some data:

[Laurent Zuijdwijk @laurent_zw]

@boxwrenchx It is a nice side effect of more predictable output. I didn't train for that specifically.