Qwen3.8-35B-A3B Spotted in GitHub Code, Optimized for Consumer GPUs
cephaloform · x · 2026-08-15
A GitHub commit in the ms-swift repository confirms the existence of the Qwen3.8 family, including a 35B total parameter model with 3B active parameters (A3B). Following the MoE recipe of its predecessor, it targets 8-16GB VRAM setups, potentially offering strong performance for consumer hardware.
More from Models
- llama.cpp adds support for Moonshot Kimi-K3 text model — pmttyji · 2026-08-15
- GLM 5.3 report reveals use of synthetic RL environments — burny_tech · 2026-08-15
- Qwen2.5 vs Qwen2 vs Gemma 4: Benchmarking 30B-class open models — MaySaki2 · 2026-08-15
- Testing Fable 5 distillation redirects and SVG generation — teortaxesTex · 2026-08-15
- Gemini 3.7 Flash Generates Entire Interactive Websites in Real Time as You Browse — _philschmid · 2026-08-15
- Testing Qwen 3.8 27B minimal thinking in a specific harness — rosie254 · 2026-08-15