Artificial intel

local ai

[Artificial intel @Artificial19601]

@ybouane Also add mcp support, so we can use agents init

[Artificial intel @Artificial19601]

@seanhighness I'm addicted to qwen 3.8 flash next on strata , but thank you qwen3.8 27b for all

[Artificial intel @Artificial19601]

@ItsmeAjayKV Also test the iq3_xxs , along with this

[Artificial intel @Artificial19601]

@Firmament_labs @coldniko @threejs @Alibaba_Qwen @QwenDevs So sys ram era officially started , no matter what the vram but with enough system ram we got local agi

[Artificial intel @Artificial19601]

@seanhighness Il try it

[Artificial intel @Artificial19601]

@satellitedown Any #rtx5090 guys have you tried this ,
How many t/s r u getting at 2lakh context at xhigh

[Artificial intel @Artificial19601]

@satellitedown I tried it in Deepseekharness, I'm getting 120-140t/s in blender mcp etc.. rarely hit 180t/s , my gpu power capped at 400watt ,

[Artificial intel @Artificial19601]

@seanhighness Ooh so I can just use spare cheap drive for this so my fast m.2 safe

[Artificial intel @Artificial19601]

@ItsmeAjayKV I try using the unsloth provided mtp heads with - ista- lab qwen varients but speed dropped from 30t/s to 19t/s ,

[Artificial intel @Artificial19601]

@seanhighness @ciruai Yup q3 is top notch , il definitely try ,

[Artificial intel @Artificial19601]

@seanhighness At default tweaks got 30t/s upto 2lakh context ,
With mem lock and parallel slots set to 2 i get 40t/s avg upto 2 lakh context,
Also tried ' Ista labs q3 ' but it dont have mtp so initial chat 30t/s and speed dropped to 20-26t/s as chat grows.
60-70t/s is needed

[Artificial intel @Artificial19601]

@seanhighness No matter what I try on rtx5090 , max I get is 40t/s with 64gb ram, I tried ud q2 k xl , it would be better if i get 50t/s average,

[Artificial intel @Artificial19601]

@PeasantSmith Which version unsloth or ista ? Which quant ? Pc specs ? T/s?

[Artificial intel @Artificial19601]

@satellitedown Keep posting benchmarks on all testings 😀 on rtx5090

[Artificial intel @Artificial19601]

@seanhighness Protect From ?
I think GOOD ASI protect us from BAD ASI ,

[Artificial intel @Artificial19601]

@DavidOndrej1 What about deep seek harness?

[Artificial intel @Artificial19601]

@WPPA @Oluwaphilemon1 Ha that's true , is there any guy who is testing open models on rtx5090 with nvfp4