HN user
from my brief search:
real-world OPNsense (yea, that's not Debian) benchmarks show that even low-power Xeon-D (Skylake-D SoC) with 10G SFP+ NICs:
pushed 9.40 Gbps in plain router mode
7 Gbps with packet filtering and no rules
7 Gbps in NAT mode
5.5 Gbps with a default ruleset
The router probably should work without relying on LLM:
I remember I misconfigured it, and can't really Google around or ask LLM to fix it, because my Internet is down now.
Of course I can use Internet on my phone, but it's not really convenient setup to ask LLM to troubleshoot the router.
About 2 years ago I tried to use Mikrotik, as the router for my ATT Fiber.
It's royal pain.
Basic router setup is not hard.
Port forward to expose web server? Can be done after heavy docs&forum reading, but hair pinning never worked for me.
I checked their WiFi AP. It's UI is hard even for senior software engineer.
In the end I got opnSense installed in VM on Proxmox and it's my router for last 2 years.
I'm using 2 Mikrotik switches with their simplified UI to operate VLANs. Zero complains.
Does Android have this feature?
And with MTP (or other speculation techniques) you can ~double that.
dual CPU is bad idea for inference: single CPU works few times faster than dual because of NUMA limitations.
where that usage data coming from?
There is no standard.
The higher quantization - the better results, but more memory is needed. Q8 is the best.
I'm getting 30 t/s on RTX 4090D (using 42 out of 48GB VRAM) with UD-Q6_K_XL
https://huggingface.co/unsloth/Qwen3.6-27B-GGUF/discussions/...
I'm running it on my Intel Xeon W5 with 256GB of DDR5 and Nvidia 72GB VRAM. Paid $7-8k for this system. Probably cost twice as much now.
Using UD-IQ4_NL quants.
Getting 13 t/s. Using it with thinking disabled.
Please offer new clients try it: at least let us to send few requests in the chat.
Thank you. Just tried it.
UI is rather confusing.
Great project!
Is there similar project for image editing?
Just basic features:
- cropping
- rotating
- brightness & contrast
So, only Americans can use data against others?
By the way, I'm running 400B model on my computer with 72GB VRAM: Qwen3.5-397B-A17B-GGUF/UD-Q4_K_XL getting 13 t/s. Subjectively, I feel it's runs at the level of Anthropic Claude, just slower.
Are you talking about Biden?
The Keystone XL pipeline had been partially constructed before President Biden revoked the permit on January 20, 2021 on his first day in office. About 300 miles had been completed when TC Energy officially abandoned the project.
They told us that with AI you can vibe-code anything now...
So, no need to make old program to work. Just write new one.
/sarcasm
The HuggingFace link is published, but not working yet: https://huggingface.co/MiniMaxAI/MiniMax-M2.1
Looks like this is 10 billion activated parameters / 230 billion in total.
So, this is biggest open model, which can be run on your own host / own hardware with somewhat decent speed. I'm getting 16 t/s on my Intel Xeon W5-3425 / DDR5-4800 / RTX4090D-48GB
And looking at the benchmark scores - it's not that far from SOTA (matches or exceeds the performance of Claude Sonnet 4.5)
That screenshot / video on README page is mostly unreadable. Can't get anything out of it.
Looks like Incus has no GUI?
Proxmox has nice web GUI
Reading few blogs and forums about it today - people talking about switching to Gateway API (from "legacy" Ingress).
And I do not understand it:
1. Ingress still works, it's not deprecated.
2. There a lot of controllers, which supports both: Gateway API and Ingress (for example Traefik)
So, how Ingress Nginx retiring related / affects switch to Gateway API?
And there is no profit, too.
Google released MedGemma model: "optimized for medical text and image comprehension".
I use it. Found it to be helpful.
I tried Talos few month ago. Found it unstable and complicated; reported few bugs.
And because they are "immutable" - I found it's significantly more complicated to use with no tangible benefits. I do not want to learn and deal declarative machine configs, learn how to create custom images with GPU drivers...
Quite a few things which I get done on Ubuntu / Debian under 60 seconds - takes me half an hour to figure out with Talos.
I'm running k3s at home on single node with local storage. Few blogs, forum, minIO.
Very easy, reliable.
Without k3s I would have use Docker, but k3s really adds important features: easier to manage network, more declarative configuration, bundled Traefik...
So, I'm convinced that quite a few people can happily and efficiently use k8s.
In the past I used other k8s distro (Harvester) which was much more complicated to use and fragile to maintain.
Proxmox has built-in support for CEPH, which is promoted as VMFS equivalent.
I don't have much experience with them, so can't tell if it's really on the same level.
That is the problem. LLMs can't be trusted.
I was searching on HuggingFace for the model which can fit on my system RAM + VRAM. And the way HuggingFace shows the models - bunch of files, showing size for each file, but doesn't show the total. I copy-pasted that page to LLM and asked to count the total. Some of LLMs counted correctly, and some - confidently gave me totally wrong number.
And that's not that complicated question.
Great idea... But appears to be not practical.
I tried that Unsloth R1 quantization on my dual Xeon Gold 5218 with 384 GB DDR4-2666 (about half of memory channels used, so not most optimal).
Type IQ2_XXS / 183GB, 16k context:
CPU only: 3 t/s (tokens per second) for PP (prompt processing) and 1.44 t/s for response.
CPU + NVIDIA RTX 70GB VRAM: 4.74 t/s for PP and 1.87 t/s for response.
I wish Unsloth produce similar quantization for DeepSeek V3, - it will be more useful, as it doesn't need reasoning tokens, so even with same t/s it will faster overall.
Russia is no more terrorist state, than USA is.
That was USA scorched Vietnam. That was USA killing civilians in Iraq and Afganistan. That was USA overthrowing foreign goverments, including Ukrainian... And then it preached to Russia on what to do with neighboring states...
2x CPU system may be slower for LLM than 1x CPU system.
Because in 2x CPU system, the model may have to be passed via NUMA, which has 10% - 30% of memory bandwidth bandwidth