Qwen3.5-397B

Headlines · 2

1. 5d
   Apple's LensVLM compresses long context into images, then zooms in

   Apple's foundation-model team and Duke University propose LensVLM, which first lets a model skim compressed page images that cover the full text, then calls a tool to expand the pages relevant to the question and read the original content. The paper reports 68.9% average accuracy across seven text QA benchmarks at 4.3x effective compression, close to the 72.4% full-text reference; under 15x initial compression the base model scores 18.1% versus 52.1% for LensVLM. Built on Qwen3.5-9B-Base, it uses Qwen3.5-397B to generate tool-use trajectories for supervised fine-tuning, followed by DAPO reinforcement learning; the work was accepted to NeurIPS 2026 and a LensVLM 9B demo is online.

   机器之心 · 🔥 0
2. 6d
   Minisforum's MS-S1 MAX-P495 AI workstation packs 192GB unified memory at $7,399

   Minisforum has opened pre-orders for the MS-S1 MAX-P495, a 3.5-liter workstation built on AMD's Ryzen AI Max+ PRO 495, offered in a single configuration with 192GB of unified memory and a 2TB SSD at $7,399 (listed at $9,249), with shipping expected in mid-October 2026. The processor change over the MS-S1 MAX launched a year ago is modest: still 16 Zen 5 cores and 32 threads, boost clock up 100 MHz to 5.2 GHz, and the NPU rising from 50 to 55 TOPS (about 10%) for a 131 TOPS platform total; the real upgrade is memory, with LPDDR5X growing from 128GB to 192GB, up to 160GB of it assignable to the GPU, and rated speed going from 8,000 to 8,533 MT/s. Minisforum says one unit ran 284B-parameter DeepSeek V4 Flash at Q4 quantization at roughly 12 tokens/s, two linked units ran Qwen3.5-397B at about 16 tokens/s, and a four-system cluster loaded a 380GB Q4_0 build of DeepSeek-R1 671B without a stated speed; all are mixture-of-experts models activating roughly 13B and 17B parameters per token.

   TechRadar · 🔥 0

Experience and discussion from the community

Share my Qwen3.5-397B experienceAsk about Qwen3.5-397B