expired
Legion Tower 7i Gen 10 (Intel) with RTX™ 5090: Ultra 9 285K, 64GB RAM, 2TB SSD $4161 $4160.69
$4,161
Lenovo
Missed This Deal? Get alerts when similar deals are posted.
LenovoDeal Details
Lenovo has Lenovo Legion Tower 7i Gen 10 Desktop (90Y80000US) on sale for $4160.69. Shipping is free.
Note: If you don't see the sale price apply eCoupon code GAMEON10 in cart. This was an expired FP deal but it's back in stock.
Processor
Intel® Core™ Ultra 9 285K Processor (E-cores up to 4.60 GHz P-cores up to 5.50 GHz)
Operating System
Windows 11 Pro 64
Graphic Card
NVIDIA® GeForce RTX™ 5090 32GB GDDR7
Memory
64 GB DDR5-5600MT/s (UDIMM)(2 x 32 GB)
Storage
2 TB SSD M.2 2280 PCIe Gen4 Performance TLC
AC Adapter / Power Supply
1200W
WIFI
Wi-Fi 6E 2x2 AX & Bluetooth® 5.3
Keyboard
USB, Calliope, Black - English (US)
Others
Pointing Device
USB Calliope Mouse (Black)
Warranty
1 Year Mail-in
Add-ons
Xbox Game Pass
Cooling System
250W 360mm Liquid Cooling + 1 x Rear + 2 x Top with ARGB Fan
Part Number: 90Y80000US
https://www.lenovo.com/us/vipmemb...90y80
Note: If you don't see the sale price apply eCoupon code GAMEON10 in cart. This was an expired FP deal but it's back in stock.
Processor
Intel® Core™ Ultra 9 285K Processor (E-cores up to 4.60 GHz P-cores up to 5.50 GHz)
Operating System
Windows 11 Pro 64
Graphic Card
NVIDIA® GeForce RTX™ 5090 32GB GDDR7
Memory
64 GB DDR5-5600MT/s (UDIMM)(2 x 32 GB)
Storage
2 TB SSD M.2 2280 PCIe Gen4 Performance TLC
AC Adapter / Power Supply
1200W
WIFI
Wi-Fi 6E 2x2 AX & Bluetooth® 5.3
Keyboard
USB, Calliope, Black - English (US)
Others
Pointing Device
USB Calliope Mouse (Black)
Warranty
1 Year Mail-in
Add-ons
Xbox Game Pass
Cooling System
250W 360mm Liquid Cooling + 1 x Rear + 2 x Top with ARGB Fan
Part Number: 90Y80000US
https://www.lenovo.com/us/vipmemb...90y80
Community Notes
About the Poster
Deal Details
Community Notes
About the Poster
Lenovo has Lenovo Legion Tower 7i Gen 10 Desktop (90Y80000US) on sale for $4160.69. Shipping is free.
Note: If you don't see the sale price apply eCoupon code GAMEON10 in cart. This was an expired FP deal but it's back in stock.
Processor
Intel® Core™ Ultra 9 285K Processor (E-cores up to 4.60 GHz P-cores up to 5.50 GHz)
Operating System
Windows 11 Pro 64
Graphic Card
NVIDIA® GeForce RTX™ 5090 32GB GDDR7
Memory
64 GB DDR5-5600MT/s (UDIMM)(2 x 32 GB)
Storage
2 TB SSD M.2 2280 PCIe Gen4 Performance TLC
AC Adapter / Power Supply
1200W
WIFI
Wi-Fi 6E 2x2 AX & Bluetooth® 5.3
Keyboard
USB, Calliope, Black - English (US)
Others
Pointing Device
USB Calliope Mouse (Black)
Warranty
1 Year Mail-in
Add-ons
Xbox Game Pass
Cooling System
250W 360mm Liquid Cooling + 1 x Rear + 2 x Top with ARGB Fan
Part Number: 90Y80000US
https://www.lenovo.com/us/vipmemb...90y80
Note: If you don't see the sale price apply eCoupon code GAMEON10 in cart. This was an expired FP deal but it's back in stock.
Processor
Intel® Core™ Ultra 9 285K Processor (E-cores up to 4.60 GHz P-cores up to 5.50 GHz)
Operating System
Windows 11 Pro 64
Graphic Card
NVIDIA® GeForce RTX™ 5090 32GB GDDR7
Memory
64 GB DDR5-5600MT/s (UDIMM)(2 x 32 GB)
Storage
2 TB SSD M.2 2280 PCIe Gen4 Performance TLC
AC Adapter / Power Supply
1200W
WIFI
Wi-Fi 6E 2x2 AX & Bluetooth® 5.3
Keyboard
USB, Calliope, Black - English (US)
Others
Pointing Device
USB Calliope Mouse (Black)
Warranty
1 Year Mail-in
Add-ons
Xbox Game Pass
Cooling System
250W 360mm Liquid Cooling + 1 x Rear + 2 x Top with ARGB Fan
Part Number: 90Y80000US
https://www.lenovo.com/us/vipmemb...90y80
Community Voting
Deal
Score
+7
Good Deal
Bad Deal
Community Reviews
This Legion Tower 7i Gen 10 is a high end RTX 5090 desktop that stands out for its strong CPU, 64GB of memory, and 2TB of storage in a prebuilt package. The extra RAM and roomy internals make it useful for video generation, local inference, VMs, and other heavy multitasking, and the NVMe slot is easy to reach without removing the GPU. Build quality and cable management are praised, and the packaging is described as very secure. The main drawbacks are stock and fulfillment hiccups, delayed delivery updates, and locked down memory settings that can make RAM upgrades unstable or slower. It feels like a capable all in one workstation if you can live with the prebuilt quirks.
Power users consolidating local coding and inference on one desktop, because the fast CPU, large RAM, and easy NVMe expansion support VMs, WSL, and request routing.
AI creators running image or video generation, because the 5090 plus 64GB or 128GB RAM can keep ComfyUI workflows moving while leaving room for VMs and other CPU tasks.
Builders who want a workstation with room to grow, because adding storage is straightforward and the system can be expanded to 128GB RAM for heavier multitasking.
Buyers comparing against a custom build, because current pricing can be hard to beat once the GPU, 64GB RAM, 2TB storage, and power supply are bundled together.
Memory upgraders who want plug-and-play XMP tuning, because added DDR5 can boot unstable or downclock to 4800 MT/s with no BIOS controls to fix it.
Users expecting a quiet, lightweight local LLM box, because concurrent GPU workloads can make inference and coding painfully slow.
People who need a machine ready on a fixed date, because delivery can slip and the order may be postponed after purchase.
Shoppers who need immediate availability, because the system can be in and out of stock and orders may be lost in shipping when inventory disappears.
Common Questions (5)
AI-generated summary based on comments from the community. Updated ago
Join The Conversation
Share your experience with the Slickdeals community
Share information with the community. Please follow our Community Guidelines and be kind!
34 Comments
Sign up for a Slickdeals account to remove this ad.
Will fix that another day. Annoying pre-built problems.
At least I got 128GB now and system ram clock down by 17% doesn't seem that critical with what I am using it seems yet.
Will fix it at some point but it is annoying if you do have this computer and upgraded the RAM and could run into the same problem.
DAMN lol that sucks
then I put in the samsung in the secondary slots then it booted
\
First it did not book then I put in my ram only at the same place there samsung was
then I put in the samsung in the secondary slots then it booted
\
I'll just move the 6400 back to the system it came from and maybe pick up an r9700 for it if I get good agentic use from models than can run on this lenovo.
I'll just move the 6400 back to the system it came from and maybe pick up an r9700 for it if I get good agentic use from models than can run on this lenovo.
my comfyui was taking up 30GB VRAM and 64GB system ram.
the extra ram now allows me to run
Linux VM (so I can have better control and automation using scripts)
I also just moved Hermes agent inside of this VM. I can give it 8GB and use the gui browser when I need it also.
The hermes agent is not accessing the 5090 local GPU for inference though, I still have online subs and I rather keep inferences going there. I found local inference on 5090 too slow and concurrent use sucks. ie: you can't expect running coding and agents against the same 5090 and not feel extreme slowness immediately. I need to find a better way to route my requests, I had a post somewhere about have to manage LLM use vs. image/video generation use because if they compete for the same GPU, the machine is very very very very slow.
Sign up for a Slickdeals account to remove this ad.
I am still at 4800, can't seem to fix it but I do enjoy the ram.
my comfyui was taking up 30GB VRAM and 64GB system ram.
the extra ram now allows me to run
Linux VM (so I can have better control and automation using scripts)
I also just moved Hermes agent inside of this VM. I can give it 8GB and use the gui browser when I need it also.
The hermes agent is not accessing the 5090 local GPU for inference though, I still have online subs and I rather keep inferences going there. I found local inference on 5090 too slow and concurrent use sucks. ie: you can't expect running coding and agents against the same 5090 and not feel extreme slowness immediately. I need to find a better way to route my requests, I had a post somewhere about have to manage LLM use vs. image/video generation use because if they compete for the same GPU, the machine is very very very very slow.
I moved hermes and etc to the Legion because the CPU is not doing much most of the time and I gave it 128GB RAM so it will not interfere on system ram even if I. am doing video. an old laptop was running hermes before. less machines to have to talk for it all to work.
https://slickdeals.net/f/19496511-lenovo-legion-pro-7i-16-2-5k-240hz-oled-intel-ultra-9-275hx-rtx-5090-32gb-ddr5-1tb-ssd-2999?p=18408938
I consolidated a few things into this 5090 because of the extra RAM and CPU cycles.
I also installed coding agent on the VM on this 5090 so I can edit and run code here and not have have my laptop open. I change my setup often depending on what feels easier. I did just install WSL Linux if I really need some Linux apps to have direct access to the CUDA cores. Again, could be a pain in the ass to coordinate requests to avoid race condition on the GPU workloads but something you have to figure out.
I worked on this request routing/prioritization based on my workflow. I do have multiple 5090's and I do split video and certain LLM requests to other machines. I just want to tweak a single machine setup first so I can keep 1 machine really humming all the time.
The watchdog script will eventually have some remote control where I can just remotely toggle what I want a machine to do and the client slide all have fail-over setup and they will just auto connect to the available LLM or ComfyUI end points.
yes, all configured with tailscale because I take my Main laptop everywhere and I wanna have access to everything no matter where I am and it works pretty damn sweet right now. I also code from my phone at times and can update certain apps which I do constantly.
I still have GLM and Minimax online sub - my hermes is on them and so is coding --- you will fine Hermes excruciatingly slow -- try it and see when you want to do anything else on the GPU, you will hate it
Because any large amount of LLM. work will bottleneck everything up also. (currency issues and some think time is just ridiculous still). Coding is too painfully slow compared to online sub so I have not made a real switch to local, slow speed is too annoying for me.
oh LLM requests, my code will try local for certain tasks and if multiple local machines are not available, (cuz I kill LM studio if video request is active) then I call free Groq for LLM as last resort. I could just configure it to go to GLM or Minimax but just never bothered to change my LLM retry config.
Some of these numbers on ROI for all this silliness.
Total cost: 23 hours of video footage generated (ai slop for sure) = 82,800 seconds ÷ 5 = 16,560 units × $0.26 per clip= $4,305.60. It's around $0.26 for each 5 seconds videos. some models cheaper but 0.26 is fine. Some are even $0.5 I was using online services and about $200 in, I realized I should just buy 5090 and pray the speed or generation and the sheer amount will be more cost effective for me. Fun to just tell people I have a few 5090's at home.
Apparently I've already AI slopped $4300 worth of online video API calls! That's one 5090 desktop already.
I think i made the right call to just buy 5090 since I anticipate ton of AI slop during my trail and error to learn things. Could be more careful in coding but sometimes I just let it rip and review a lot later and just keep playing with the code/settings/prompts/workflow. Fire off a few things during the day, do my work and review outcome afterwards. I also code from my phone since I created coding vm on the 5090 so it's always on and I don't always need to claude code it from my laptop.
I just love that I can learn things constantly as I can talk to claude all day to figure stuff out.
long post again but I love to hear what you are trying to do so I can learn new stuff too!!!
I upgraded to 128GB system because My CPU was doing nothing so I added ram so I can run apps that are CPU bound to make more use of the machine and now I have a very very fast VM running coding (LLM is the slowest part) and agents. and I can run a lot more CPU tasks when I get a chance to add more.
For $700 more, I get to use 24 CPUS on this Legion so I upgraded to 128GB ram for $700 Seemed like a good trade at the time.
I couldn't run VM without the extra RAM, it was competing with the video generation (idk, maybe it was the embedding model ,vae ertc, some of those models get swapped in and out in Comfyui so you def want min 32 system ram and 64gb appeared to be the sweet spot but then you're wasting the available CPU cycles.
I may actually run another LLM inferencer for small models and LOCK it to CPU only and get decent enough token generations. I can do less LLM vs. ComfyUI requests routings then is my thinking.
optimize optimize optimize.
BTW, if you are running LTX, there are workflows that uses q4 gguf that can run as low with 8GB VRAM but it was pretty slow and a loot more nodes to configure. It was buggy for me so I killed it. Now I just let it rip and consume the RAM it wants on the default workflows and not try the low vram workflows. WAN2GP supposed to be easier and more efficient but I haven't made a real switch to it yet.
should be poweful enough
I will always use Mac as primary.
I got the Omen first (it has less CPU) but I was going after the GPU anway and now I have to sell it or buy more ram cuz it only has the AMD 8 core cpu (~9700?) and 32GB system ram.
Now I got the Legion, the CPU is faster (not as fast as M5) but got 128GB and can always run so I started to shift some development right on the Legion so it is always on and I can mobile connect to my claude session for cheap.
I thought I would use 128GB Macbook with local coding LLM but it is just way too slow and ineffective. Now I have a very nice MBP M5 Max 128GB that I run maybe 35B moes but random stuff but NOT as a coding LLM server. I did started to run some VMs because it's very fast. (I don't really do dockers but I got a few images running). VMs will use up my CPUs a lot more than I can and GPUs are not used as much as I thought I would.
$100 year MiniMax token plan will smoke any coding llm I run on the Mac, speed, smart, and # of prompts you can throw at it. (I have this plan on top of my Pro GLM plan -- before the price hike).
I tried to use it as my prompting engine a quick comparison with my 5090, generating a long series of video prompts for some reason took 40 minutes on M5 Max (lol! - a lot of think time) but 5090 did it in 6 minutes. I gave up trying to do anything fast + meaningful LLM on the Mac if one of the 5090 is available.
Mac is too smart sometimes, I tether often and a first few times I was excited an hooked up a lot of jobs to the M5 for LLM use, the machine churned thru the full 10% in about 60-90 minutes because the LLM was running so and hot LOL!!! The machine was running even though it was unplugged and closed in my backpack!!! The macbook bottom was hot af.
Now I don't have anything wired against the M5 unless it's like 4th choice. LOL
Just beware if you are using any online subscription LLM today and then you go to a single Mac (512GB even as it does not matter), you will HATE the speed and slowness you get. abd there is no more 2nd session, no sub-agent capabilties cuz you will be stuck in a single file for your Mac to respond.
5090 running the best Qwen 27B coder is slow. I forgot the exact but the ram requirement for 27B dense is much higher than 35B moe. M5 Max was 30 tgs best case on my. If use oMLX or dlash or mtp, still not gonna be fast enough and unable to parallel process requests. I think some people got m5 max to go 40 tgs 27B. There were a lot of benchmarks but I tried it and it was still not good enough. 600 GB/s is just too slow when 5090 is 1700+GB/s (but less VRAM made it not that good with 27B.
35B is faster and less ram hungry but it's not as smart. 5090 was 200+ tg/s while M5 max was 100 tgs/s. You can get more if you squeeze dflash/mtp but you can't fix parallelism problem.
Hermes / claw are pigs on context so it will burn your GPU cycles like crazy if they start spamming requests, you will see your inference engine queuing up requests and can't even processing prompts fast enough.
If people code full time on local LLMs, they must really have secrets , have super patience, or have no idea how fast online LLMs are cuz I simply can't stand the slowness.
I need to draw a diagram of my setup but I keep changing so I just forced myself to try to consolidate what I think should be on the Legion since that's the fastest "overall server" now and see how much I can squeeze out of one machine and go from there.
I am in the process of installing Ollama on Legion with CPU-only inferencing so I can use up more system ram and cpu without taking any GPU. I think the CPU is fast enough to get like 20-30 tgs on 8B models but I need to double check. This frees up a lot of ram on 5090 and 8B will run a lot faster as a secondary meta prompter if the 5090 GPU is queued up for ComfyUI. Lots of weird logic I have so far with all these machinary.
Sign up for a Slickdeals account to remove this ad.
Join The Conversation
Share your experience with the Slickdeals community
Share information with the community. Please follow our Community Guidelines and be kind!