rapidTools / Mike

[rapidTools / Mike @Rapidtools99525]

@ayayalar @MiaAI_lab You can use it with the previous recipe. You just have to modify the env and it works perfectly fine. I get approx 40-60+ tps.

[rapidTools / Mike @Rapidtools99525]

@MiaAI_lab @huggingface Alright. What's your tps? I get approx 40-50 right now with your latest recipe. Seems weird why... Drafter is 66 tps and 40-50 acceptance rate.

[rapidTools / Mike @Rapidtools99525]

@MiaAI_lab @pradeepXkapoor This is something that I will believe only when I am running it on my machine. If it's true, I will delete all the other models. 😂

[rapidTools / Mike @Rapidtools99525]

@MiaAI_lab Mine was 77 so far. But it's really fast. Ibwish Mimo would be this fast...

[rapidTools / Mike @Rapidtools99525]

@XMonetizationC_ LOL... BS as it's best... This is what it can do in real: https://t.co/jVFqI2ZwEL

[rapidTools / Mike @Rapidtools99525]

@MiaAI_lab Actually it's not just smarter but handles long context better and can solve problems with less thinking, especially when it's using web search.

[rapidTools / Mike @Rapidtools99525]

@MiaAI_lab Yeah for sure. I get approx 17-30 currently. It's usable.

More would be pretty awesome, so I support the idea. ;)

[rapidTools / Mike @Rapidtools99525]

@MiaAI_lab @thatcofffeeguy Soketimes I have. Probably the use case is different. (3dsmax sdk and other low level macro heavy stuffs)

For that it's not always that reliable unfortunately. Mimo handle those cases better I think.

[rapidTools / Mike @Rapidtools99525]

@MiaAI_lab What I noticed that it's weirdly slow especially for an A3B model.

I am using the NVFP4 quant with DFlash. Reducing the mtp to 8 and also the temperature make it a bit faster, but it's still somewhere close to 30 tps.

[rapidTools / Mike @Rapidtools99525]

@SkillsMastery_ Just heat it up and you can push together the two pieces. These are very bad solutions. Tubes should be watertight.

[rapidTools / Mike @Rapidtools99525]

@jun_song MiMo v2.5 is somewhere between Hy3 and DS4F. It's pretty reliable for C++ and math.

[rapidTools / Mike @Rapidtools99525]

@MiaAI_lab @aleksandar_xyz Yeah but the quality difference is significant. Actually I don't what language / framework / sdk / api. For me it's 3ds max mostly and it has a pretty big and macro heavy code base. In that case Mimo gives noticable better results.

[rapidTools / Mike @Rapidtools99525]

@aleksandar_xyz @MiaAI_lab It's much better in planning. The code quality seems more structured. It keeps the context much better and degrades much less over time as the context gets bigger.

[rapidTools / Mike @Rapidtools99525]

@JustJoeFX @MiaAI_lab Yeah that's whatvthey say, but honestly it's pretty damn slow. After longer context drops below 8 t/s...

[rapidTools / Mike @Rapidtools99525]

@MiaAI_lab Mimo is doing the same thing. Actually it's a good thing because it's not going to make BS and the chancen of having an actial working solution is much higher.

[rapidTools / Mike @Rapidtools99525]

@Bhavani_00007 That's IGI level graphics...

[rapidTools / Mike @Rapidtools99525]

@MiaAI_lab @NVIDIAAI Hmmm. Surprisingly fast. Multimodal and 'theoretically' better than DS4Flash...

[rapidTools / Mike @Rapidtools99525]

@MiaAI_lab I have started downloading it 30 min ago... I have to check this onassive c++ projects. I am seriously curious about it.