The consensus right now is that Qwen3.6 in its 27B and 35B-A3B versions is better for coding whereas Gemma4 is stronger when it comes to OCR, audio transcription and the likes. Margins are slim though and the harness at these model sizes is the most important factor.
I can see that and I don't know your setup, but there are people pushing >70t/s with MTP on a single 3090, with big contexts still >50t/s. 64k is not a lot for agentic coding, and IIRC 128k with turboquant and the likes should be possible for you. r/LocalLLM/ and r/LocalLLaMA/ are worth a visit IMO.
"[...] In the coming days, we will also open-source smaller-scale variants, reaffirming our commitment to accessibility and community-driven innovation. [...]"
It‘s mainly due to system requirements that Flux.2-dev doesn’t get same usage as Z-Image. A 5090 needs about a minute to generate an image with a basic workflow with Flux.2-dev. But prompt adherence and scene/character consistency in edit mode is (way) ahead of Qwen-Edit-2509 if you ask me.
I feel tired reading about him and it doesn’t even phase me anymore. It‘s just another thing I add on to the pile. Maybe it’s part of the plan to go numb to everything he does.
I mean they make a bold statement up top just to paddle back a little bit further down with: "[…] In terms of Chinese and Cantonese recognition, the SenseVoice-Small model has advantages."
The Guardian, especially for their podcasts, is the only news website I am paying and have ever payed for. And I pay more willingly than any other newspaper would get from me for their paywall stuff. It‘s that valuable for me to support this approach.