Dawid Babula

Developer · creator of Eternal Vows & Zenith · building on DGX Spark · Ai and Crypto

[Dawid Babula @Hiceron2]

@MiaAI_lab @MiaAI_lab got 43 but needed to make 65k draft-vocab file for it ( cant find yours in repo ;) ) great job. i can feel the speed difference already

[Dawid Babula @Hiceron2]

@sudoingX hah hello neighbour Spark enthusiast XD happened to me last week in Rayong with slightly bigger rock. let us know how does 3.8 next and 5.3 flies on 2 sparks, i still didn’t both my second unit...

[Dawid Babula @Hiceron2]

@ModelScope2022 "Before release of Full family of qwen 4 " -thats a strong hint that qwen 4 family of models will still be released as open source/open weights @QwenDevs 🫰

[Dawid Babula @Hiceron2]

@bullbear_info @MiaAI_lab cap your Mhz settings to 2200-2000 , put spark sideways and invest 5 $ in usb fan. i use 27b with h3 and trellis2 (3d modeling ) and am under 80C ;)

[Dawid Babula @Hiceron2]

@Tech2Wild @Alibaba_Qwen Yes I talked about it too😁 if we get same size and quatize to 2bit for dwarfstar or 3bit mia solution... man can dream🤣

[Dawid Babula @Hiceron2]

@mr_r0b0t @NVIDIAAI does it have vision or you use something for vision?

[Dawid Babula @Hiceron2]

@Shreeyutm05 @RexFan11 @xieenze_jr Thats amazing can you point us to source of your information? I cant find it

[Dawid Babula @Hiceron2]

@MiaAI_lab @MiaAI_lab does non thinking 47 tps is worth more then thinking 29 tps? Every time I see comparisons I go with no, but every time I see posts from anyone showing off the potential speeds I see structural. How is your experiance comparing both Mia ?

[Dawid Babula @Hiceron2]

@0xWhiteMage sounds amazing, hope you succeed and let us know when can we expect to test it

[Dawid Babula @Hiceron2]

@Blackwellboy it looks like dgx is somehow bricked at gpu level and ai can suggest that too. don’t be fooled :)

[Dawid Babula @Hiceron2]

@Blackwellboy 1/2 there could be other reason (happened to me before) and it's caused by failed update via dgx dashboard. if you ever have similar drop in clock power and hard unplug doesn’t work you will need to load back previous working version and if that fails do it via bios.

[Dawid Babula @Hiceron2]

@jun_song @MrTacticalX You know what would be cool? 2bit obliterated Super version of dsv4 flash 0731 that could run on one spark with dwarfstar and @MiaAI_lab recipe ;)

[Dawid Babula @Hiceron2]

@renegademtl @LocalInference @jun_song if you get some info about place with lower prices in Thailand i would love to know XD

[Dawid Babula @Hiceron2]

@LocalInference @jun_song that’s so bad, was planning to buy second unit and they up the price. i both mine for 116 000 bath in february, wish i took 2 of them at once xD can we have the ram price crash already?

[Dawid Babula @Hiceron2]

@mr_r0b0t @MinLiBuilds of course. spreading the good work ;) i do a lot optimisation but you , Mia and few others pioneering it hard! great job

[Dawid Babula @Hiceron2]

@MinLiBuilds try 0.7 gpu uti with dspark2 recipe from @mr_r0b0t https://t.co/GrxUgOaFYd
it gives 29 tps in prose already and over 30-40 in coding thinking mode. using 4 agents on xhigh thinking with over 60 tps (need to measure again, had over 60 on 3 agents). its doing great work :)

[Dawid Babula @Hiceron2]

@Vladimir76897 @MinLiBuilds steady 25-29 (dflash2) or 24 for sglang in prose. 1000 tps prefill speed. for coding its faster. thinking 35- 50 no think can benchmax to 75. i use thinking mode. dont understand the reason for using this model without thinking from results shown. hope it helped @Vladimir76897

[Dawid Babula @Hiceron2]

@MiaAI_lab @shuai_bai_ @Alibaba_Qwen It would be funnier if we get comparable size to deepseek v4 flash and got possibility to use dwarfstar to run it like we do dsv4 sladh 2bit on one spark

[Dawid Babula @Hiceron2]

@bleysg Great job @bleysg do you plan to adjust it for glm 5.3 when they release open weights on 2 sparks and add qwen3.8 family ?