nacs

With Strata and specialized engines, Qwen Flash Next runs even on 16GB cards.

I use it at 60+ tok/second on a single 24GB card.

They might have spent next to 0 time actually 'hacking' - they may have just told Opencode to hack Opencode website and let it run for hours or days till it succeeded.

Hm? Jev's actual (network) latency is not that great and even small local models are doing far better latency-wise.

The Microsoft article GP linked even shows the MS model having 95ms latency.

Even on price Jev being matched (the same MS model is "Input tokens cost $0.042 USD per million tokens. Output tokens are free.", same as Jev).

Cloudflare's Clef-flash model is actually slightly cheaper: "$0.038 in / $0 out per 1M" too.

Jev is being matched or exceeded in performance and price within a month of them going public.

Coder is not as good and seems to have noticable loss.

Use the normal one - I've used that for long coding sessions and it worked well on multi-hour runs.

@the_mitsuhiko - How is the session replay of Pi embedded in the article?

It doesn't look to be asciinema?

I'm seeing 128GB DDR5 RAM costing $3500+ just by itself.

How are you going to build an entire system shipped for $3k?

This spec sheet looks more like a dreamed-up spec sheet than pricing reality.

Is the Magic 8 ball immune to hallucination as Jev is?

It would be good to list 1) zero-shot accuracy and 2) latency on the models page . The LLM-based models' latency is probably much higher than the BERT approaches I would assume.

Also curious, it seems from looking at the accuracy scores you gave that it seems to be NLI > Gliclass > Laya (for Bert types)? Why do you seem to feature/recommend Laya more - is Laya better in some way?

Apple product won't be the cheapest but it is a full package (CPU, RAM, VRAM/GPU, fast-storage, etc).

If you look at the pricing of a full (x86) AI workstation you'd need around the nvidia GPU, you'd approach $10k easily (and be using a ton more wattage too).

Good to know thanks.

That's with CPU offload to a DDR5 6000 RAM though which is around $3-4k at least.

5090 has 32GB VRAM.

It'd be silly to buy the 18k model to run a tiny model like Qwen 27B. You use models like GLM Flash and Qwen Next which won't fit on a single 5090.

People don't buy Sparks and M5 Ultras to run a 27B model - you buy it to run an MoE model like Qwen Next which this M5 excelled at.

That's a dense model. Of course it will do worse.

Now try running that Qwen 3.8 Next model on the 5090 and tell me what TPS you get (hint: it's near 0 since it doesnt fit the 32GB VRAM on 5090 vs the 256 in OPs M5).

It was yes, lots of words in the input to describe state of the game-world at each checkpoint.

Easily. It took around 4GB VRAM on my nvidia card so a 48GB unified mem should have no issue. < 1s inference

Tailscale funnel works well but you can only have 1 per Tailscale instance (1 per machine) at the https endpoint so if you want to have a few apps, the others will go on non-https, custom-port URLs.

Intended to be publicly accessible internet content, I have no problem with.

But are you ignoring the literal scanning (and 'burning down' of books) that Anthropic has been found guilty of? Or the torrenting of pirated content en-masse by Meta that there is an active lawsuit over to name just 2 recent examples?

Look at the image and audio/video models especially - they can reproduce everything from Mickey Mouse (the copyrighted one) to making entire Seinfeld episodes with the real cast (both visual likeness and even the actor's voices).

Claude/OpenAI etc have taken and continue to take literally all data from the internet, printed books, image, audio, and video humans have created in all of existence without permission to train their models.

But when same AI company gets "distilled" or it's own AI-generated content used to train other models, it's suddenly immoral or illegal?

Openrouter lets you pin or blacklist providers or specify provider per-request as well.

Did you miss the part of this article where the benchmarks specifically call out Fireworks as one of the worst in their tests:

> Fireworks scored 46% on TAU, a 30 point gap

Another surprise was DigitalOcean being bottom of barrel too.

Companies are apparently willing to risk their brand name by being deceptive about these heavily quantized/flawed model-serving.

And of course it just renders an HTML document (applies some CSS to the text content) and then uses the embedded browser's printing.

I gave it another go with the hint above about the tetherd log:

  tail -f ~/.local/state/tether/tetherd.log

I killed that persistent process (sudo killall), rm /run/user/1000/tether/tetherd.* (to clear the lock file), restarted tetherd after pairing to phone via Linux bluetooth settings.

I was then able to see and pair over wifi and bluetooth. Everything is enabled except Notifications (even though both address book sync + notifications are enabled in iOS bluetooth settings).

Clipboard sync working now as well as send/receive files.

The tether app still has occasional issues re-connecting to tether though when I move in and out of the app.

I ran into similar issues on a mostly stock Ubuntu 24.

When I start the gtk app, it has a scan devices button but it never finds my phone even though I've paired via bluetooth and its on same LAN.

The iOS app just has a spinning icon that goes forever even though it sees my computer name in the 'discovered devices' list.

I tried pairing from CLI and GTK app and no go on Wifi or Bluetooth.

I thought so too and scrolled the page looking for a last name. Then realized the ending was not the same. RIP

What's the point in having it use an offline LLM if it still requires an internet connection to send all your chat messages through the "Groupmind" paid service?

Seems like the real purpose of this project was to plug Groupmind and whatever else the author made.

He had thinking mode set to "low" fortunately.

It's looks like they literally cloned the UI and graphics from the Mini Metro/Mini Motorways games (not just the map style).

The vast majority of companies are still putting most things in the cloud and will continue to do so and this far outnumbers the must-be-on-premise companies.

Sure there will be self-hosters but hosting AI models will always be more of a challenge than running scalable database on your own hardware and specialized hyperscalers will be here.

Reminder that by using Bitlocker, you're using a closed source encryption for which Microsoft will happily hand out your recovery key on request.

https://www.forbes.com/sites/thomasbrewster/2026/01/22/micro...

UBlock Origin is also giving me the warning on that domain.