@antirez, @dcapitella c'è ancora molto lavoro da fare ma nel frattempo A 256K i risultati sono **AR 1214 PP / 25,88 TG** e **MTP 1196 PP / 59,91 TG t/s** su qwen 3.8 q2 antirez-recipe
@ivanfioravanti I'd be more worried about people who don't know how to open a bottle, but Citroën is the one bringing real innovation with a 15,000-euro electric car. Fiat bought Ferrari by selling the 500, and Volkswagen bought Lamborghini and Bentley by selling the Beetle.
@antirez , @dcapitella Preliminary test: results for antirez/qwen3.8-flash-next-gguf on Strix Halo without mtp, obtained with an LLM I'm building by combining ds4, gufo and tensorfold. I think they should carry over to ds4 too.
[🔺#GIVEAWAY🔺] À l'occasion de la sortie officielle de CONTROL Resonant avec #RTXON, tentez de GAGNER 1x carte graphique NVIDIA GeForce RTX 5080 customisée aux couleurs du jeu !
Comment PARTICIPER ?
🟢 RT + Commentez avec #RTXON
@antirez@FrameworkPuter Good news—just when I've managed to get the Qwen 3.8 Flash prefill up to nearly 1,700 t/, AMD could be the game-changer thanks to its still-affordable prices.
We're giving away a CONTROL Resonant custom wrapped NVIDIA GeForce RTX 5080 to celebrate its official release with #RTXON.
To enter:
🟢 Share this post
🟢 Comment #RTXON
T&Cs: https://t.co/Mbf7w3j1q3
@ivanfioravanti@thybead semplicemente.... i progetti che puntano all'AI locale si contano sulle dita di una mano, l'altra parte del mondo sta spingendo per vincolarci sempre di piu alle AI di frontiera e comincio a trovarlo sempre piu sospetto.
@antirez@cloverfield198 There are too many greedy people in the world to hope for a scenario like that; I expect more of a monopoly like the one with blood diamonds.
@antirez I had created a branch with Strix Halo + DGX Spark; I reached 79 t/s in prefill and 4 t/s in decode over LAN for GLM 5.2. If it helps, I can open a pull request—I still need to deploy it to GitHub.