Jacob Verdoorn

591 posts

Jacob Verdoorn banner
Jacob Verdoorn

Jacob Verdoorn

@VerdoornJacob

Coordination of Investment, Tax, and Estate Plans for CEO, Founder, and Multi-Gen Families | Client Advisor | 7 person team | 100 years combined experience

Minnesota Katılım Ağustos 2022
506 Takip Edilen229 Takipçiler
Jacob Verdoorn
Jacob Verdoorn@VerdoornJacob·
@sudoingX Thanks for all your help... I'm not the expert at this. I just point Grok or GPT at your work and it helps give ideas on how to tune based on all the work you have done... Thanks for telling people to buy up cards a few months back... $600 3090 feels good now... Cheers
English
1
0
8
394
Sudo su
Sudo su@sudoingX·
this is exactly what i needed, thank you. saving these flags. 100 tok/s on hermes with q4_0 kv cache and flash attention, that's the real world config right there. running the same setup on the 5090 when 3.6 lands on my beast, will tag you with the comparison. curious if the 5090 mobile hits similar thermal walls at sustained load
English
3
0
12
1.3K
Sudo su
Sudo su@sudoingX·
85-100 tok/s on the 3090 with qwen 3.6 already? that's in line with what 3.5 MoE was doing. drop your full flags and context length you tested at, i'm pulling 3.6 on the 5090 24gb and will run the same config for a direct comparison. if anyone else is running qwen 3.6 on a 3090 or any consumer card drop your tok/s, quant, and flags below. building the community benchmark sheet before i publish my own numbers
Jacob Verdoorn@VerdoornJacob

@sudoingX 3090 getting 85-100 t/s on cpp server with new qwen3.6 35b a3b ud q4 k m 262k context

English
45
2
266
20.6K
Jacob Verdoorn
Jacob Verdoorn@VerdoornJacob·
@sudoingX This is also running at 260w on the 3090 because the card is temperamental.... and will black screen me if pushed to hard...
English
0
0
1
120
Jacob Verdoorn
Jacob Verdoorn@VerdoornJacob·
@sudoingX Qwen3.6-35B-A3B-UD-Q4_K_M.gguf \ -fa on \ -ngl auto \ -c 262144 \ -ctk q4_0 \ -ctv q4_0 \ -kvo \ -fit on \ -fitt 1024 \ -fitc 4096 \ -t 8 \ -tb 8 \ --chat-template-kwargs '{"enable_thinking": true}' Getting 100t/s using Hermes Best local experience yet…
English
6
2
43
1.5K
Jacob Verdoorn
Jacob Verdoorn@VerdoornJacob·
@sudoingX I like the speed of 3.6 3090 getting 85-100 t/s on cpp server with new qwen3.6 35b a3b ud q4 k m 262k context now to see if it can build...
English
0
0
1
105
Sudo su
Sudo su@sudoingX·
right now qwen 3.5-27b dense Q4_K_M is the king on single 3090, 35 tok/s flat, one-shotted octopus invaders with 1,827 lines. the 3.5 MoE (35B-A3B) was faster at 112 tok/s but different tradeoffs. testing 3.6 this week on the 5090, same 24gb vram class so the numbers will translate directly to your 3090. stay tuned
English
1
0
2
358
Sudo su
Sudo su@sudoingX·
i spent the entire afternoon at a cafe downloading 6 models to a 5090 laptop for the next wave of benchmarks, walked in the door, sat down, and qwen 3.6 just dropped while i was pulling qwen 3.5 open source does not let you breathe. 35B total, 3B active, agentic coding matching models 10x its size, multimodal, thinking mode, apache 2.0. this is the successor to the exact model that hit 112 tok/s on a 3090 and buit octopus invaders in 14 minutes with 3,483 lines of code the 3.5 MoE was already the fastest model i ever tested on consumer hardware. if 3.6 improves on that with better agentic coding at the same efficiency, nothing else in this weight class comes close pulling it on the 5090. qwen 3.5-27b dense, carnice-27b, gemma 4, and now qwen 3.6, all loaded on the same 24gb machine. the comparison data is about to be disgusting. speed sweeps and octopus invaders dropping soon
Qwen@Alibaba_Qwen

⚡ Meet Qwen3.6-35B-A3B:Now Open-Source!🚀🚀 A sparse MoE model, 35B total params, 3B active. Apache 2.0 license. 🔥 Agentic coding on par with models 10x its active size 📷 Strong multimodal perception and reasoning ability 🧠 Multimodal thinking + non-thinking modes Efficient. Powerful. Versatile. Try it now👇 Blog:qwen.ai/blog?id=qwen3.… Qwen Studio:chat.qwen.ai HuggingFace:huggingface.co/Qwen/Qwen3.6-3… ModelScope:modelscope.cn/models/Qwen/Qw… API(‘Qwen3.6-Flash’ on Model Studio):Coming soon~ Stay tuned

English
48
18
595
40.5K
Jacob Verdoorn
Jacob Verdoorn@VerdoornJacob·
@sudoingX 3090 getting 85-100 t/s on cpp server with new qwen3.6 35b a3b ud q4 k m 262k context
English
10
0
22
21.5K
Sudo su
Sudo su@sudoingX·
let me clear something up for the new followers. the 5090 mobile has 24gb vram, same class as the 3090. when i benchmark a model on the 5090 and give you the flags and the tok/s, that translates directly to your 3090 at home. the architecture is newer so the 5090 numbers will be slightly faster maybe, but the configs are identical. if it fits on my machine it fits on yours. and i'm not stopping at one gpu. 3090 nodes are still in the rotation for controlled comparisons, smaller gpus are coming for the 8gb and 12gb crowd, and nvidia sent me a dgx spark that's clearing customs right now. 128gb unified memory on my desk soon. 7 models loaded on the 5090 today, hermes agent work i've been cooking for weeks is almost ready to ship, and open source keeps dropping new models faster than i can pull them. the benchmark pipeline is about to run nonstop. i am so soo back.
English
18
4
135
8.5K
Jacob Verdoorn
Jacob Verdoorn@VerdoornJacob·
@sudoingX Curious what you find best for the 3090. I am loading up 3.6 for my 3090
English
1
0
1
284
Sudo su
Sudo su@sudoingX·
lets see how far this 5090 goes before i melt these rog keys. 7 models loaded on one machine, 113gb of weights, 1.3tb still free, and open source keeps shipping faster than i can benchmark qwen 3.6 vs qwen 3.5 dense vs carnice-27b vs gemma 4 dense vs gemma 4 moe. all on same hardware, same quant, same harness, no excuses from any of them keeping you all posted. data first, opinions after
English
4
1
44
7.3K
Jacob Verdoorn
Jacob Verdoorn@VerdoornJacob·
This is amazing. I was literally talking to an old book dealer at a conference school conference this weekend about something very similar. My idea was to digitize all of these old out of print books that may not be on the Internet yet to preserve them and use them in curriculum
English
1
0
1
69
Jesse Genet
Jesse Genet@jessegenet·
Passion project alert 💕 thechildslibrary.com My @openclaw agents and I built this (free!) resource for filling your child’s bookshelf with time tested books you are sure to treasure Fave feature: filter by “values” 💛
English
96
63
961
61.4K
Jacob Verdoorn
Jacob Verdoorn@VerdoornJacob·
@jessegenet @edlndn @openclaw @daylightco We are doing the same reading book you are... and would love this... or i'm just sending my claws to try to copy it lol Thank you for sharing all of your ideas and journey. It is very exciting and inspiring for my wife and I as we are developing homeschool for our kids!
English
0
0
1
75
Jesse Genet
Jesse Genet@jessegenet·
@edlndn @openclaw @daylightco I want this, so far everything I’ve built has a level of customization and bugs I know about that feel a bit too rough to ship without instructions but I want to get there soon and just share so people don’t double up on my efforts
English
4
0
12
831
Jesse Genet
Jesse Genet@jessegenet·
Built my dream phonics / reading app using my @openclaw team for my @daylightco e-ink device Now reading lessons can happen outdoors, total game changer for my kids moods on tough days ⛰️🌷
English
27
17
502
61.4K
Jacob Verdoorn
Jacob Verdoorn@VerdoornJacob·
@cordes_tax keychron.com/products/keych… I think I have this one… I love the volume knob It types so smooth. And I bought silent switches, so I can type while on phone calls. (Many custom / gamer keyboards have very loud clicking keys…)
English
0
0
1
107
Matthew J. Cordes, EA
Matthew J. Cordes, EA@cordes_tax·
Best keyboard for Mac users that also use Windows on Parallels? Everything I see keeps saying the Logitech MX Keys S. Anyone have this one?
English
14
0
6
1.5K
Brian C. Beasley
Brian C. Beasley@BrianCBeasley·
Imagine a hypothetical “straight-line” investment existed—one that delivered perfectly steady, constant returns every single year and matched the exact long-term performance of SPY (the S&P 500 ETF) since 1993. (Of course, no such investment actually exists.) Even with that perfectly smooth ride, many investors would still have fired it multiple times over those 33 years—simply because it looked like it was “lagging” during certain stretches. Can you spot when most people would have caved and switched to SPY from the straight line? And how badly would that decision have hurt their long-term results?
Brian C. Beasley tweet media
English
5
0
9
589
Brian C. Beasley
Brian C. Beasley@BrianCBeasley·
I've been an advisor for 31 years. Most of you need to burn the following ten thingsinto your soul: 1. Know your goals. 2. Fund them, using conservative assumptions. This is huge. 3. Invest systematically, and with rigid discipline. No emotional bets allowed. Have a system for "flyers" if you must. But your system should allow for you being blatantly wrong on those. You must win regardless. 4. As you approach withdrawal mode, manage drawdowns and overall volatility. This kills if you get it wrong. 5. Be aware of taxes. Minimize, but don't be blinded by them. A lot of people "throw the baby out with rhe bathwater" just to save on some taxes. 6. Insure against things that would be disasterous. 7. Keep your beneficiaries, estate plans, and legal documents up to date. 8. and THINK LONGER TERM. Your fickle strategy switching and recent performance chasing will set you back. 9. At some point, you or your spouse will want to violate one or more of these. People who listen to good advisors will never violate these. 10. Ignore or fail to meet with good advisors at your own peril.
English
2
0
9
195