Remixie

@Bushwina iq3_xxs is too low quality of a quant. tune your params to run q4_k_m

Everyone is boasting about their qwen 3.6-27b mpt success , but running small quant at low context what's the point? I tried to run hermes agent on a single 3090 iq3_xxs to fit the min 64000 context. Not worth it. #rtx3090 #locallma #hermesagent

64k context length

Qwopus3.6-35B-A3B-v1-Q5_K_M.gguf LM Studio 32 GB of DDR5 + the 32 GB VRAM of a 3090+3070 But it flies at around 50-70t/s with a normal level of tool use Is it smart? Not as much as 3.6 27B from my initial testing in opencode but boy is it faster

I could technically get more t/s if I used vanilla llama.cpp but I’m tired of searching for flags that actually exist

@outsource_ You don’t need to do all this. Just use the nvfp4 version of 27b

Benchmarking Qwen3.6-27B on a 4090👇🏻 Found the clean recipe: mainline llama.cpp + Qwen3.5-4B same-vocab draft → 43 tps mean / 67 peak, 6/6 quality pass at 32K context. Also found that the popular "150 tps" cross-vocab spec-dec recipes silently corrupt JSON output. Per-category bench tells the truth. Configs, raw run logs, dead ends (DFlash OOM, TQ3_4S, vLLM cudagraph), and the bench harness: Next: training a same-vocab Qwen3.6-Drafter-0.6B for 100+ tps clean. GitHub with all runs and data 👇🏻 github.com/outsourc-e/qwe…

@GiaaOW I’m sorry for whatever is happening

dude. fuck cancer, actually.

@frank_vega_cl I’m renting a windows vm with these specs. It’s nice to have the freedom to run whatever we want for cheap nowadays.

@Remixie if I could only get a 5090 in banana country

Guys. Forget everything you’ve heard. Get Qwen 3.6 27B NVFP4 running on llama.cpp on a 5090 with 48gb of ram to get 50t/s+ with 128k context length for actual decent code editing in opencode. It’s wild that this is possible now.

@meta_alex Tune your params. See the press release footnotes for SWE benchmarks

@Remixie What does your config look like I’ve tried on a rtx6000 pro and very unreliable tool calling

AND IT HAS VISION

@chiefofautism No. Just use whatever model is available for free or subsidized. Or mess with models on rented machines. These two tactics have dissuaded me from buying hardware for AI.

Hand posing practice

@hebitsukai 😄

迷路描きました。どうしようもなく暇な時にどうぞ。

@hebitsukai Every day that hebitsukai posts an illustration, it is a good day 😎

Tried my hand at the bee challenge over at vgen. I can't shade to save my life but at least I'm slowly improving my line art

It's crazy how magical art fundamentals are. I can still barely draw and yet I was able to do this by just drawing the base first.

@albertadevs If you ever get acquired by OpenAI, they will achieve ASI through memery

Practice some Loomis method today #animeart

@xSaandi Are you interested in getting help with your site? I've done website work for others in the streaming preset market

Tomorrow ill be posting a video on subtitles. If you dont know how to use subtitles presets, this will be perfect for you! 😁

@mulemount Ecuador?

mas latinitas