Qwen 3.5

Headlines · 2

1. 21h
   PrismML releases Ternary Bonsai 2 on-device models

   PrismML on Oct. 15 announced the Ternary Bonsai 2 series of on-device models in 2B, 4B and 9B sizes, built with ternary weights that encode parameters as only -1, 0 or +1 and take up roughly 0.43GB, 0.95GB and 2GB of storage respectively, joining the previously released 27B version to cover lightweight to complex local workloads. The company says the models target local agents on phones, PCs and wearables, handling instruction following, everyday tasks, tool calling and image understanding; across 19 benchmarks spanning six capability areas, the 9B model used about 40% fewer tokens on standard instruction execution and everyday math tasks. PrismML claims up to 5x throughput and 3x energy efficiency versus equivalent full-precision models on hardware such as the RTX 4090 and M5 Pro, with the 9B reaching up to 34 tokens per second and the 2B up to 120 tokens per second on an iPhone 18 Pro Max; the models are built on Alibaba's Qwen 3.5 base, released under Apache 2.0, and runnable via llama.cpp on NVIDIA GPUs or MLX on Mac, iPhone and iPad.

   IT之家 AI · 🔥 17
2. 2d
   Surface Laptop Ultra hands-on: Nvidia RTX Spark brings local AI to Windows on Arm

   On Oct 7 Microsoft held a Windows and Surface event in San Francisco with Jensen Huang and Satya Nadella on stage, opening preorders for the RTX Spark N1X-powered Surface Laptop Ultra and the first wave of RTX Spark OEM machines, which ship Oct 16. The top configuration pairs a 20-core Grace CPU, a 6144-CUDA-core Blackwell GPU and up to 128GB of unified memory, with Nvidia claiming up to 1 PFLOPS of AI compute at FP4 with sparsity. In Ifanr's hands-on tests, Qwen3.5 35B-A3B (Q4) ran at about 59.65 tokens/s and Qwen3.8 27B at about 20.98 tokens/s locally under Ollama + Open WebUI, WorkBuddy could download local models and run agent tasks offline, and pricing starts at 21,988 yuan with the 128GB version at 48,388 yuan.

   爱范儿 · 🔥 3

Experience and discussion from the community

Share my Qwen 3.5 experienceAsk about Qwen 3.5