SenseTime

Why we recommend SenseTime SenseNova

SenseTime (商汤科技), founded in 2014, is one of China's largest AI companies and listed on the Hong Kong Stock Exchange in 2021. Its foundation-model platform "日日新 / SenseNova" is built around native multimodal — unlike "text-first, vision-bolted-on" stacks, SenseNova unifies understanding and generation in a single architecture from the ground up.

SenseTime uses multiple domains. The brand portal at www.sensenova.cn is where you read about the SenseNova models and the Token Plan — the "Visit official site" button at the top points here. The developer surface at platform.sensenova.cn is split into two halves: /console is where you mint API keys and view billing, and /docs is where you read the API reference + OpenAI SDK integration guide. The actual API base URL is https://token.sensenova.cn/v1. All three subdomains run on the same SenseTime account — sign up once and you can use all of them.

SenseNova Token Plan is the personal-developer subscription, currently in public beta — the Free tier (¥0/month) is available now with limited capacity; Lite / Pro paid tiers are coming soon.

The Free tier fully includes SenseNova 6.8 Flash Lite (a lightweight multimodal agent — 64K / 65,536 tokens max output, with upgraded multimodal understanding and data analysis, improved across 6 benchmarks) + SenseNova U1 Fast (an accelerated U1 variant on the NEO-Unify architecture, focused on infographics and high-density PPT generation) — both at 1,500 calls / 5 hours. DeepSeek V4 Flash (1M context / 64K output) and the newly listed GLM-5.2 (Zhipu's flagship long-context text model — 1M lossless context / 128K max output) are also available in the same list, on a tighter 500 calls / 5 hours quota each.

The model number moved from 6.7 to 6.8. Per the official docs, a compatibility route is enabled through August 31: calls to sensenova-6.7-flash-lite are automatically redirected to sensenova-6.8-flash-lite. Switch to the new model ID after that.

On performance, 60% token savings — on long-chain tasks like information search and deep research, native-multimodal SenseNova averages 60% fewer tokens than pure-text agent models, with significantly more deliverables (charts, PPTs, reports, infographics) per unit cost. That's why "long tasks can run with confidence."

On ecosystem, Cowork-Skills is an 8-component stack spanning understanding / execution / generation: material analysis, multi-source retrieval, table understanding + image analysis → data analysis conclusions, PPT interactive refinement → PPT generation, report writing, infographic generation. Skills are freely combinable by industry — education, academic, finance, retail, manufacturing, gov, healthcare and more. The platform integrates with Hermes Agent and OpenClaw out of the box; the docs also ship an Anthropic-compatible endpoint plus setup guides for Raccoon, Claude Code and Cursor, so existing toolchains switch over at almost zero cost.

The open-source side is active too — model weights are public at github.com/OpenSenseNova (SenseNova-U1 / Vision / SI / MARS / Piccolo Embedding / NEO / Kairos / SenseNova-Skills). The flagship SenseNova U1 / U1 Pro targets production image generation (with an interleaved image-text chain-of-thought, quality matching top overseas models) and is currently billed outside the Token Plan quota as a higher-tier pay-as-you-go capability.

If your project needs native multimodal capability (image + text + infographic) + long-chain Agent execution + Chinese office-scene specialization + a friendly 1,500 calls / 5 hours free quota during the public beta + reliable China-region access, SenseNova Token Plan is a 2026 pick worth reserving a spot on now — once Lite / Pro go live, you're first in line to migrate into production.