Qwen3-TTS does excellent cloning and is very fast. On my workstation (5060Ti) it renders the audio at around 2.5x realtime. I’ve added an example with my parameters now.

I’ve added example scripts for client, server and MCP configs. It should be much easier to replicate with those now.

Oh thanks all for the nice comments! My fork is at https://github.com/troed/speech-to-speech but beware it’s not kept in any other state than what I’ve last pushed :) Also you’ll have to supply your own chime wavs and voice sample etc.

When I start llama-server I point it to the models-config that have unique max context sizes per model - and they’re allocated at their max size as soon as the server starts so since it comes up I will be able to use that context size too.

I’m actually a bit unsure as to how you run it since you get OOMs during usage :)

I also use the DCP plugin for Opencode to help manage the context cache and have less of a disruption as it gets compressed, but I wouldn’t need to for the above to work. When I hit the context limit the context would still get compressed.

Less-than-beautiful UI added to the clients (still works as CLI if run with parameters). Somewhat problematic finding an easily cross-compileable Go UI lib.

Also made the typical pipe-to-bash installation flow, awaiting proper ujust in some far away future.

Regular 1Gbit/s with three switches in total between them - distance about 40m (different buildings).

Jag har hittills haft åsikten att vad jag än kan välja för bredband så väljer jag (och rekommenderar andra) Bahnhof eftersom oavsett om de är billigast eller inte så vet jag att mina pengar går till något bra för oss alla.

Med Telenor som ägare kommer jag nu istället gå strikt på features vs kostnad. Antingen levererar de bäst eller inte.

I think this hadn’t been made a big deal out of if it hadn’t been for the conspiracy theories on Proton just before

Yeah it was quite obvious from all the “clanker” that this is a typical run-of-the-mill AI-hater, but I found this funny anyway:

The weight count is too low to reproduce the training set verbatim

That’s not what anyone wants. That’s not how LLMs gain “intelligence”, at all. The whole point of training on large datasets is to NOT internalize training data verbatim.

Oleksandr Kravchenko, a partner at U.S. consulting firm McKinsey & Company, will be Ukraine’s economy and ecology minister

o_O

The trick in itself (the FFN tensors selectively being off GPU) shouldn’t be.

Not sure I understand but I’m on Linux fwiw.

Yeah I run the Gemma 4 12B QAT on my homeserver. It has a 12GB GPU (Nvidia A2000) and that model gives the most bang for the buck being both fast and multimodal.

My daughter loves the Moana movies. I’ve told her about this one - and the anti-excitement of “the same movie but with actual persons” is quite obvious. I can only conclude that no actual parents were involved in its inception.

Our framework outpaces orbital velocity by 9 times on a consumer GPU

I’m sure that nomenclature is obvious with their intended audience but am I confused right now

Sounds more negative than it is. +8% in Europe. The Orange buffoon’s trade war and China’s anti-competitive internal practices aren’t much to do about.

Lemmy and Mbin are roughly the same thing - “Reddit”

Mastodon is “Twitter”

… and there are more, with Loops being “TikTok” and Pixelfed being “Instagram”.

They can all see the same content, but presented in ways that emphasize their unique aspects. That’s what people call “Fediverse” but it’s much simpler to just think of them as separate apps and ignore that technically true but confusing aspect.

Kan inte läsa materialet på något sätt som skulle innebära att Mastodon eller Lemmy skulle inkluderas. Faller både på att “små företag” exkluderas samt “civilsamhällets organisering”.

IANAL

Yeah I’m more thinking about the four wheel drive Gen 4.

They’ll never be allowed to compete head to head. It’s like Formula E, that will always be enough different from Formula 1 so as to never show what a pure matchup would look like.