SenseNova 6.8 Flash tops agent benchmarks vs Opus 5, reads 582-page filings in 8 minutes
智东西 · wechat · 2026-11-03
SenseTime released SenseNova 6.8 Flash, claiming top scores on TeamBench and DataAnalysis 2.0 ahead of Claude Opus 5, GPT-6 Sol and Kimi K3, plus 93.33 on LongMemEval-S and a joint first on BEAM 128K.
In hands-on tests via the Xiaohuanxiong desktop agent, it compared two 582-page Anthropic IPO filings, correctly extracted 6 updated figures and built an interactive dashboard in 8m21s; PPT generation took 64 minutes mostly on formatting. It also supports dynamic posters, agentic 3D in Blender, SVG animation, and Figma integration.
SenseTime recently open-sourced SenseNova Loop-9B-Preview, a 9B looped-transformer, pairing closed-source products with open-source architecture research.
More from Models
- Claude Power User Burns the Midnight Oil as Weekly Opus Quota Drops to 6% — rudrank · 2026-11-04
- DeepMind study: a heart emoji flips Claude & GPT from 0% to 100% recursive code — dioscuri · 2026-11-04
- Reddit blocks all LLMs except Gemini, leaving devs stuck in a data loop — banteg · 2026-11-04
- Microsoft's 4B Mage-VL-Cowork hits 66% on OSWorld-Verified SOTA; Niels Rogge builds runnable demo on HF Sandboxes — NielsRogge · 2026-11-04
- Microsoft's 4B Mage-VL-Cowork sets on-device computer-use SOTA, demoed fully on HF — NielsRogge · 2026-11-04
- Manor Labs walks back AI self-improvement hype after closer inspection — manorlaboratory · 2026-11-04