HN user
thbl AT fastmail.com
A lot of replies here are about Mac devices and their support for these 27B models. I own a MacBook but use a Lenovo Thinkstation PGX to run my models. It has a gb10 Blackwell gpu and 128gb unified memory. You can connect multiple ones.
If this is real I will stop my monthly subscriptions.
Finally! Now it becomes economic for us to make the move! Goodbye CloudFlare!
As a Nvidia Spark owner I can confirm that Linux is much more performant then Windows on that chip. If the goal is "doing stuff with AI" Windows is the last OS you want to look at.
Isn't voxtral much better?
Mostly ecommerce and pricing data. I work for marketplaces, brands, retail stores and even our own saas competitors. We match the EAN (gtin) to the correct SKU within seconds (Google Shopping, Amazon, etc). Part of it is our own trained ML models.
Despite that I find the goal of what you are trying to achieve questionable, I believe it should not be the AI that judges you here.
We are all witnessing the start of an AI era that will not end soon. Guiderails are a part in this development. I do have questions about the people, or systems, that decide on what's good and bad behavior. This tech is used in any country in the world. As long as they are able to pay their subscription in dollars, someone is able to use it. Is it up to a company to decide what's good or bad behavior? Is this a debate? Is this politics? Is this just a vision of one company? Would it shift in time? Will it be stricter for more hyper-intelligent models? Will it change when open source models are becoming better and better?
With scraper tech I mean a rust binary that is able to download and process thousands concurrent urls (millions per hour). Not to the same domain obviously. Paying more is not the issue here, its more the idea that an AI decides on what part of the spectrum I operate. Why is it opinionated? I am not doing anything wrong, why does it make me feel like I have to defend myself.
I am not using brew. Just checked and it still says 2.1.98. Will try manual update.
I am on the latest version available for me 2.1.98
I built a distributed DuckDB setup using OpenRaft for state replication. Every node holds a full copy of the database. Writes go through Raft consensus, reads are local. It's more like etcd-with-DuckDB than MotherDuck-lite.
OpenDuck takes a different approach with query federation with a gateway that splits execution across local and remote workers. My use case requires every node to serve reads independently with zero network latency, and to keep running if other nodes go down.
The PostgreSQL dependency for metadata feels heavy. Now you're operating two database systems instead of one. In my setup DuckDB stores both the Raft log and the application data, so there's a single storage engine to reason about.
Not saying my approach is universally better. If you need to query across datasets that don't fit on a single machine, OpenDuck's architecture makes more sense. But if you want replicated state with strong consistency, Raft + DuckDB works very well.
I'm building a Redis real-time backup platform with PITR (point in time recovery) to the exact millisecond. Besides downloads it has an emergency recovery option where you connect the Redis client directly to the master-replica to receive the latest version.
The key explorer let's you change data on the fly and receive notifications in real time when a condition is met (if value contains X).
It's build in Rust on bare metal wit isolation between clients and data.
ReplicaSafe.com (nothing there yet, will take a few weeks)
I think this model proves it's very efficient and accurate.
I am on Linux for 26 years. Last 5 years I run PopOs! on my desktop. User friendly and stable, Ubuntu based.
China and Europe (Mistral) show that models can be very good and much smaller then the current Chatgpt's/Claudes from this world. The US models are still the best, but for how long? And at what cost? It's great to work daily with Claude Code, but how realistic is it that they keep this lead.
This is a new tech where I don't see a big future role for US tech. They blocked chips, so China built their own. They blocked the machines (ASML) so China built their own.
I am the co-creator of ShoppingScraper. Convert an EAN / GTIN to pricing information, product specs, content or image. API-based and rapid with pricing and barcode data.
The website needs some love, but the webapp is going well.
They say: Dashboard Login Latency and Timeouts
They mean: You server will get connection timeouts through CloudFlare proxy
We hire a developer to build parsers for a complicated file format. It takes a week per parser. Gemini 3 is the first LLM that is able to create a parser from scratch, and it does it very well. Within a minute, 1-shot-right. I am blown away.
I use Gemini cli, Claude Code and Codex daily. If I present the same bug to all 3, Gemini often is the one missing a part of the solution or drawing the wrong conclusion. I am curious for G3.
How does it handle questions asked by Claude?
I use Comet since today and I believe that can do exactly what you need. I ordered stuff on Amazon autonomous.
They’ve built performance, enterprise utility, privacy, sovereignty, open innovation and strategic partnerships into their core story. It's quite a list. The models are opensource, Voxtral outperforms Whisper in terms of accuracy.
There is no AI company like Mistral.
I use it on a 24gb gpu Tesla P40. Very happy with the result.
Pretty cool, but where is the Android app?
If I talk to a business relation or friend, they give me an answer with a few words or a few sentences. Asking a question to an LLM gives me paragraphs. I am not sure if a different interface is going to change that.
I use LLMs now by asking questions, copy/pasting stuff around. A human-feeling in terface sounds more like TTS and I don't think that is human for me.
About the Webapp; better search and filter on previous conversations. Filters on model type. Better errors when context is too big. Forking conversations would be nice. Better export options. Copy whole convo (not just response or reply).
On the LLM: It's too positive. I don't always want it to follow my ideas and I don't want to hear how much my feedback is appreciated. Act like a machine. Also the safety controls are too sensitive sometimes. Rlly annoying because there is no way to continue the conversation. I like gpt4.5 because i can edit the canvas. Would like to have that with all models.
Also some stats like sentiment and fact check would be nice. Because it gives nuances in answers I want to see with the stats how far from the truth or bias I am.
And the writing.. Exaggerating, too many words, spelling mistakes in European languages.
Qwen3-30A-A3B-2507 is much faster on my machine compared to gpt-oss-20B. This leaderboard does not reflect that.
The model is good and runs fine but if you want to be blown away again try Qwen3-30A-A3B-2507. It's 6gb bigger but the response is comparable or better and much faster to run. Gpt-oss-20B gives me 6 tok/sec while Qwen3 gives me 37 tok/sec. Qwen3 is not a reasoning model tho.
It's not in my interest that a tech company from the US decides what a bad actor is.
It was fun because it was open. Now it's just another brand seeking dollars.