Momus - Local AI Lab
Region: US
Sunday 18 October 2026 15:50:53 GMT
25461
1425
105
125
Music
Download
Comments
Nick Matteo :
What inference engine are you using you need to try out strata for flash next it’s a game changer for me it was unusable on llama cpp 300 prefill. But with strata I get 70tps 1200 prefill. With a 3090 and 64gbs ram and the quality difference from the 27b is very noticeable I would target a fine tune or even a QLORA for flash next if you better decomp
2026-10-20 20:18:45
0
simply.seymour :
Fine tune and rent data center gpus and make that a video
2026-10-19 23:40:22
0
Gavin Brown :
I have access to a computing cluster with quad a100 40gb nodes connected via nvlink. Get in touch and we can set something up.
2026-10-18 18:05:38
21
evanpantano70 :
how are you running 16 sessions co-currently on a 5090 with qwen 27b?
2026-10-19 15:04:01
0
Tony Stark :
Finaly a new video about local ai
2026-10-18 20:18:14
3
user133567223 :
Noob question but how can you run multiple qwen3.8 27Bs on 1 5090 I thought it could only hold 1 instance since it was a 27B parameter model and it would take up 27gb of vram
2026-10-19 03:18:56
1
devonmarkus :
Hey I have dual Blackwell 6000s if you need help
2026-10-18 17:27:40
6
Bunny Bytes :
keep finetuning and publish to huggingface. There's only one finetune for decompiling for Qwen3.5 and its just for C code. It'd be huge.
2026-10-18 17:06:15
20
KB UBUNTU :
Local ia is a joke the level of cash you have to spend for half bake model os sad 🤣🤣
2026-10-18 22:16:04
0
nopenope :
nice work. I want to train my model to do that. do you use lama factory to train? your training materal is the work the frontier models did?
2026-10-18 20:07:21
1
David :
27b on a 5090 and 16 parallel? what are your flags? that is crazy, I know I have dual 3060s but I am at 200mb left on each after q8 vision and q4km of the swift 1.5 27b with 64k context at fp16/fp16
2026-10-18 16:25:17
1
omg hilarious 😂 :
thank you for doing interesting stuff with your equipment finally 😆
2026-10-18 21:16:12
5
James :
So it can’t do the hard stuff. ?
2026-10-18 21:29:59
0
Not So Friendly Local AI Dev :
what tps is your flash next? are you using strata? is your fine tuned 27b open source?
2026-10-18 22:21:48
0
CHRISTIAN FLORES :
what does fine tuning mean and what’s the point
2026-10-19 20:20:48
0
Christian :
so so impressive! youre one of my sources of motivation to build a similar server in a few years
2026-10-18 17:13:55
1
kriticalkat :
Did you end up training your own instruct version or just fine tune the harness with lessons?
2026-10-18 20:13:02
1
brianpulliam2 :
Need a GitHub so we can copy your fine tuning protocol
2026-10-19 00:11:00
0
Ashen :
hell yea can i have this
2026-10-18 20:19:41
2
Armin Hartmann :
I may be misunderstanding something here, but it sounds like you're training the model on the exact code base it's supposed to work on. It sounds a bit like giving a model the answers to a test and then asking it to do the same test. In other words, would your model also outperform larger models on a different set of questions or a different codebase?
2026-10-19 13:33:25
1
Mika :
When qwen4 27b release will you finetune this to?
2026-10-18 22:10:10
2
Loertu :
Bro, will u publish it on huggingface
2026-10-19 00:43:40
0
Piz :
Very interesting
2026-10-18 16:31:52
3
GuamReeferShark :
I have a missile on my computer right now. I wiped all the weights so just like no data set on it anymore. Crazy part. I still haven’t trained the base weights, but it’s speak English now. I will not be as crazy to say that I’m not using the weights because they have a purpose , but they’re all set as they are right. And after an accidental experiment, I decided I’m completely eliminating. QLORA. I I had zero drift or hallucinations. Until we trained those. And then it came back crazy. You should look at how I designed the weight though these are live and the structure is very real, I’m using a tracer with my PC
2026-10-19 18:27:02
0
moneyprintergobrrrr :
dual dgx army
2026-10-18 17:24:16
1
To see more videos from user @momus.ai, please go to the Tikwm
homepage.
