- The architecture preview : hybrid linear/sparse attention, gated residual streams,…

This is a AI post classified by Jev as Open & local models (a model release), kept by the AI Radar because it carries real work, not commentary.

Qwen 4 is coming soon : - The architecture preview : hybrid linear/sparse attention, gated residual streams, N-gram tables that can sit in system RAM, Muon training. - The lineup being discussed: - Qwen4-Max - Qwen4-Flash - Qwen4-Plus - Qwen4-27B - The 27B tier is the one most people will actually run locally. Alibaba confirmed Qwen 4 is currently in training. The 5–10 trillion parameter plans belong to later Qwen 4.5 and Qwen 5 models, not this generation.

Posted by Md Ismail Šojal 🕷️ (56k followers) 1 h ago · 8 likes · 632 views · view the original post on X. Kept by the AI Radar as Open & local models.

More AI work like this

• I had Opus 5.5 build a tiny Sims town where every resident's dialogue is written live by… — @Tech2Wild
• Meet Qwen3.8-27B iMatrix NVFP4 MTP GGUF: a multimodal beast that reads images AND text.… — @HuggingModels
• Meet Lanni-ni/hard_2gram_pile_2layer, a text-generation transformer with a twist: it's a… — @HuggingModels
• Xiaomi have distilled MiMo V2.6 into Qwen3.5-9B and have introduced an MTP head into it… — @Biggest
• Introducing Hotdog-27B, a specialized bipartite classifier. — @michael_chomsky
• update on the same bundle. a new prefill kernel landed in the prismml fork this morning,… — @sudoingX
• We got new image gen model releases from @AntLingAGI — @ItsmeAjayKV
• Meet a 100M parameter transformer trained on the BabyLM dataset. It is a text generation… — @HuggingModels

Every post is read and classified by Jev (TypeSafe): what it is, which market it belongs to, and whether the link is a real tool. 43.2k posts from 5k X accounts over the last 14 days, 1.8k tools, 19 markets. Collected every 5 minutes, fully re-ranked every hour — last update 2026-09-22 21:33 UTC. Full method.