vr8vr8

[vr8vr8 @vr8vr8]

@yournewdad_2 @TeksEdge 2 x GB10. Nvidia dgx spark.

[vr8vr8 @vr8vr8]

@Marcelo51310919 @AndrewDev8383 rtx spark costs like 3/5 of rtx 5090 and is full thing. with 128 GB not 32 GB.

[vr8vr8 @vr8vr8]

@AndrewDev8383 it lacks QSFP56 connection. that card alone costs $1000. So it makes sense that price is lower. But you cant connect 4 of those new ones.

[vr8vr8 @vr8vr8]

@stevibe Try Hcompany/Holo-3.1-35B-A3B-NVFP4 it works better then qwen 3.6

[vr8vr8 @vr8vr8]

@spark_arena @NodeHello Realy good model. Hallucinating less then Qwen 3.6

[vr8vr8 @vr8vr8]

@0xSero So 3 x GB10 and you can run it?

[vr8vr8 @vr8vr8]

@loktar00 @dunik_7 Maybe we need to start reporting those bots for fake misleading info? To get them all down as it really sucks to see spam in timeline

[vr8vr8 @vr8vr8]

deepseek-ai/DeepSeek-V4-Flash leads in speed (30 vs 20 tokens/s) thinking and quality.

[vr8vr8 @vr8vr8]

Tested out stepfun-ai/Step-3.7-Flash-NVFP4 and results are following: model failed to make floating clouds while even every Qwen 3.6 / MiniMax 2.7 / DeepSeek V4 Flash models managed to do that properly every time. Animals look terrible. But they are grounded jumping in place not like many models get them in 0 x 0 coordinate jumping. But thinking is longest of all models 70% of all output was thinking. I thought MiniMax2.7 was a big thinker so this model beats it quit hard 🤣 In overall going back to DeepSeek V4 Flash.

[vr8vr8 @vr8vr8]

stepfun-ai/Step-3.7-Flash-FP8 looks bit better. But still not thinks way to much.

[vr8vr8 @vr8vr8]

@SpaceTimeViking As always you try to brake the rules 😂

[vr8vr8 @vr8vr8]

@LottoLabs yeap bots are storming local llm... and i see so many of them lately.... and key factor is they all say spark is 3k 🤣

[vr8vr8 @vr8vr8]

@mr_r0b0t @Snixtp checked, left some comments 😂 Good idea to educate people pining some initial setups once you start local llm jungle

[vr8vr8 @vr8vr8]

@mr_r0b0t @King0James0 Gemma-4 introduced they mtp like 1 or 2 weeks ago would be nice if you would update recipes with mtp :) plus i suspect vllm/vllm-openai:cu130-nightly can be outdated as they just recently released version 22 and that one you mention was from 19

[vr8vr8 @vr8vr8]

@Blau34 @mr_r0b0t @NVIDIAAI @huggingface i'm using just mtp=3 and i bet you use like 5? :) as acceptance rate is very high for 3 ~80%

[vr8vr8 @vr8vr8]

@mr_r0b0t @Snixtp That keyword SM120/121 haunting all rtx 5xxx and gb10 owners 🤣 and i fall for both of them...

[vr8vr8 @vr8vr8]

@sudoingX Sad part it thinks way to much compared to competitors. But nice to see new models coming and pushing competition https://t.co/tqkQXlD4hQ

[vr8vr8 @vr8vr8]

Tested out stepfun-ai/Step-3.7-Flash-NVFP4 and results are following: model failed to make floating clouds while even every Qwen 3.6 / MiniMax 2.7 / DeepSeek V4 Flash models managed to do that properly every time. Animals look terrible. But they are grounded jumping in place not like many models get them in 0 x 0 coordinate jumping. But thinking is longest of all models 70% of all output was thinking. I thought MiniMax2.7 was a big thinker so this model beats it quit hard 🤣 In overall going back to DeepSeek V4 Flash.

[vr8vr8 @vr8vr8]

@TeksEdge yeap bots are posting insane barging about prices that are already close to x2. Now w can confirm that we are at peak of local LLM?

[vr8vr8 @vr8vr8]

@mr_r0b0t @Snixtp NVFP4 doesn't have mtp support. So it wont 🤣 learned it hard way 🤣

[vr8vr8 @vr8vr8]

@atomic_chat_hq Now name the fact that it supports just tool format it recognizes and if you ask for other it will fail drastically ignoring new format. And Bonsai beats it hard in normal tests having just fraction of memory usage. https://t.co/ev9t1JGCB8

[vr8vr8 @vr8vr8]

@liquidai Just tested it out and could say prism-ml/Bonsai-8B-gguf beats this model like 10 times. Results are so bad that it can't even generate solar system lineup... But speed is insane.