Hacker Newsnew | past | comments | ask | show | jobs | submit | SamInTheShell's commentslogin

@q4 is definitely smarter than sonnet from what I’ve seen so far. It’s even caught problems in code made by fable, when using it as a code reviewer.

Read the license again.

The license you receive when you download Gemma off of Google's website is not necessarily the same license that AT&T gets when they deploy Gemma as a customer service bot. That's the whole point. AT&T can work directly with Google for a licensing and legal framework that provides certainty.

From what I’ve been seeing, the Mac studios do look like they have potential. I was looking to drop $10k-$15k on one until recently. After comparing a Radeon 7900 XTX vs Ryzen Halos 128GB vs M1 MacBook Pro 64Gb, I landed on just getting an external closure setup with Nvidia RTX 5090.

The model I’m specifically targeting to use at high speeds is Qwen 3.8 27b @q4ks. This model actually proved to be good at coding (it sits somewhere between Sonnet 5 and Opus 5 capability). M1 got 10 tok/s, Ryzen Halo 20tok/s, and Radeon 7900 XTX 50tok/s (can only do 128k context window in Radeon card).

The prefill gets extremely slow around 50k tokens in context window (whatever prompt processing stage entails could be wrong about phases here). It takes about 2 hours to fill the context.

Even with a drafter model intended for speed instead of mtp, I can’t get past 70tok/s, still is extremely slow to process prompts as context grows, and drops down to 40-50tok/s anyway making this config still moot for improvement on my Radeon card.

The only thing I can point to slowing me down is bandwidth of the card itself.

I am waiting to actually get my 5090 right now and I am betting that the 1700 Gbps of capacity will fix my prompt processing speeds. I don’t need full PCIe lane bandwidth to serve my house I just need to load the full model into vRAM and let the GPU do its thing.

Additional benefit to the external enclosure route is being able to migrate the inference between devices more easily. I can develop out the infrastructure then migrate the card to be hooked up to a shared node in the house with all the tools necessary for my family to take advantage of the privacy enhancement that comes with local inference.


How are you actually using the local model? I've played with Qwen 3.8 27b on ollama and the coding harnesses (Claude Code and OpenCode) seem to fail way more often then using the cloud models. And by fail, I mean the edits don't apply cleanly, it goes to add python code, but doesn't indent it properly, or the edit doesn't apply and so it tries again and again and eventually wipes out a different function then it "intended". It just gets really frustrating compared to the relative stability of Claude Cloud.

My use-case is only coding, every model sucks at writing good literature and there is no way around that (have had people try to debate me on this, but it's a taste thing, I have extensive English writing skills from my school years).

Prior to two weeks ago, I was just using Pi and Ollama.

I have tried my hand at putting together a few harnesses and I finally landed on what I like. Been working on this small app to handle running llama-server for me from any device that has the llama-cpp stack setup: https://github.com/SamInTheShell/loom

Qwen 3.8 is the first model I've been using that hasn't been having issues doing edit calls. Here are my llama server settings and GUFF that I use: https://gist.github.com/SamInTheShell/0bf838e8dc5093583b688e...


try using pi harness, hae not encountered these sort of problem myself also yuou can ask codex to look at the transcript and figure out the solutions to tool call failures that way

I should have mentioned in my prior comment, I didn't have any issues in Pi either.

try ninfer once you get your 5090 https://github.com/Neroued/ninfer

Just wanted to share, I had a SaaS AI drop me a script to bench ninfer against llama-cpp and it is impressive. The place that it's doing better at than llama-cpp seems to really be late in the context window.

Initial results boiled down as follows.

# lmstudio-community/qwen3.8-27b@q4_k_m decode falloff 104.3 tok/s @ 12,683 -> 55.8 tok/s @ 240,755 (53% retained) prefill falloff 3,274 tok/s -> 1,059 tok/s (32% retained)

# qwen3_8_27b_nvfp4.ninfer decode falloff 173.3 tok/s @ 11,867 -> 139.3 tok/s @ 225,710 (80% retained) prefill falloff 8,726 tok/s -> 2,816 tok/s (32% retained)

I should still have room for more performance on the table. I've not even touched the overclock settings on the GPU.

This is a really cool project, I'm going to have to get into what those 3 guys are doing... assuming it can be done with what I got.


Thanks for sharing, I'm definitely trying it out after I get through my project milestones for the 5090. The README claims 700tok/s for Qwen 3.8 27b, that would be amazing, I'm only expecting an increase from my ~50tok/s on my Radeon to 200tok/s on the 5090.

Poor software choice for usecase. `sudo pacman -S podman` didn't come with these problems out of the box and assumed rootless by default.

Junk. Just another company that doesn't understand why I pick up a phone or use a support chat or send an email. Before I even started in corporate, I stopped checking email. Now when I call in to get support for things, everytime I get a clanker, I'm guaranteed to talk to a human customer service person to chew them out and cancel the service.

This is pretty good to know today. I’m currently using cosine similarity in some of my code that could have just been dot products.


Just use a harness that discards all except the most recent thoughts. This model is probably the first real small model that does well on long horizon tasks.


I did. Distributed too and supports erasure coding. The reason is because nothing quite had the architecture or features I wanted in a deployment pattern that matched my use case. All major databases that are validated are complicated to operate. I built something that operates one way for a non business use. It’s great, doesn’t need to serve a million people, just about 5 or so. Doesn’t need to be the most optimized thing ever.

I’m opting out of any corporate targeted code for personal use, anything not designed for user experience gets scrapped when I’m the user and I have Software on Demand as a Service.


It's worth running, even quantized. I liked Meta Muse Glimmer's outputs, but qwen3.8-27b@q4_k_s kinda seems way better. Haiku/Sonnet kinda pairing in workflows?


I'm kinda in the camp of wanting nothing to do with node if I'm building in Go. We have our own stdlibs for serving.


I am on the Java/.NET camp for the most part, however I do share the sentiment.


I’m in the other camp. I switched to node backends from go because of vite.


Same, hard pass on remote shell as a service that is Node and its ecosystem.


I'm not sure I follow. Node is not on the production path


It's easy to follow, it's a knee jerk tribalism reaction and not a logical or informed one. Go attracts backend or anti-frontend people so anything intermingling the two will draw out the commenters who bemoan anything that realistically builds a modern website. Just ignore it, your choice for dev tooling is practical and makes sense to me.


As someone who has had to clean up systems that only did a _npm install_ to get compromised, despite being good Renovate citizens, it doesn't matter if it is not "on the production path".

I don't care to hear arguments about how npm was _configured wrong_.

A tool with insecure defaults, with a long and storied history of being the vector by which data has been exfiltrated and companies hacked, needs to do an awful lot to justify still being used for any infrastructure our team is responsible for.


The proxy caches and package solution for Go has had its own drama.

My big contention with this project is that if I wanted to use node (or npm; or vite) in any capacity, I would have chosen node for that. I don’t want to use those tools in this context because it serves me worse than a well thought out solution.

The thought out solution is to isolate your frontend into its own directory. You can embed and serve that as part of your Go server. That frontend could have been done with react, vite, jsx, whatever typescript nonsense you want; without the Go parts necessarily needing npm or node in the build process.

I literally have toy projects illustrating the idea (this one is a toy I had AI make with some of my own patterns months ago): https://github.com/SamInTheShell/social/tree/main/frontend

Projects like grafana do stuff like this already. It’s a known good pattern.


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: