Weight forensics with Claude: Qwen3.5-27B appears grown from 9B via layer looping
CatAstro_Piyush · x · 2026-10-29
Jonathan Chang used Claude to analyze Qwen weight relationships and found Qwen3.5-27B was grown from Qwen3.5-9B via layer looping, independently corroborated by Grok. An accompanying chart maps weight links across released Qwen models by generation, distinguishing growing (stacking/width tiling), same-shape continued training, permuted links, and duplicated layer blocks — covering dense lines across generations, MoE checkpoints up to 2.4T-A95B, and unreleased nodes like Qwen3.6, Qwen3.8 and Flash-Next (60B est.). Unconfirmed by Alibaba.
More from Models
- Users report Qwen3.8-Flash-Next produces verbose, dense, metaphor-laden docs — walden42 · 2026-10-29
- Mistral's 1T-parameter Le Chonk lands on Hugging Face with llama.cpp support — Disastrous-Work-1632 · 2026-10-29
- Mistral open-sources Le Chonk, tops FrontierCode among open models for agentic coding — MistralAI · 2026-10-29
- Meta and Oxford's OverSeer lifts multi-agent world-model consistency on Minecraft to 60.2% from 0–12.5% baselines — armand_ruiz · 2026-10-29
- Intel spinout Articul8, valued at $500M, launches multimodal decision model Bivia — rohanpaul_ai · 2026-10-29
- Mistral Large 4, the 1T-parameter 'Le Chonk', is now open weight — kuberwastaken · 2026-10-29