PixelShipper

@PixelShipper

Building with AI in public. Shipping, learning, repeating and improving. 🚀

Netherlands

Joined March 2026

Following

Followers

Replies Reposts Media

PixelShipper
@PixelShipper
Aug 16
The new AI text watermarks are probably going to make a lot of industries happy, and perhaps keep jobs relevant that were thought to be on the edge of AI replacement. Anybody with great SEO gains because of AI content is not going to be loving this though.
PixelShipper
@PixelShipper
Aug 15
If you haven't already tried this, you definitely should. This level of performance on a 27B model is incredible and a great showcase of what is to come for local inference this year.
Qwen
@Alibaba_Qwen
Aug 14
We promised open weights for Qwen3.8. Now, time to meet them! 🎉 ⚡ Qwen3.8-27B: - A native multimodal dense model. With just 27B parameters, it outperforms Qwen3.7-Plus overall and shines in real-world coding & office workflows. - 262K native context, easily extendable to 1M
PixelShipper
@PixelShipper
Aug 12
Meta is back in the open weight game, and this time it's Apache 2.0. Muse Glimmer went up on the Hub on August 10 from Meta Superintelligence Lab: a 30B dense causal transformer with a perception encoder, built for agentic work on hardware you already own. The license is the
00:00
1
PixelShipper
@PixelShipper
Aug 4
The largest open model ever released now runs on a single consumer GPU. It needs 3.72 GB of VRAM. AirLLM v3.1.0 landed on July 29 with Kimi K3 support. The author measured it end to end on one RTX 6000 Ada, against the full 1.56 TB checkpoint, generating real tokens: • Peak
00:00
1
PixelShipper
@PixelShipper
Aug 4
Qwen is open-sourcing a 27B next week. That's the one that runs on your own hardware. Alibaba confirmed open weights for two Qwen 3.8 checkpoints: the 2.4T Max, and a 27B. The 2.4T is the headline. The 27B is the one you'll actually download. Why that size keeps mattering: •
00:00
1