SenseTime SenseNova 6.8 Flash

Overview

SenseNova 6.8 Flash is a production model SenseTime released on 2 November 2026. The official positioning is a next-generation multimodal agent model, with the work concentrated in two places: carrying an office task all the way to a delivered result, and pushing design output from viewable to usable.

On the office side the keywords are long-term memory and proactive interaction. The two cases SenseTime put forward show the intent. In the first, a customer once turned down a partnership because the product lacked private deployment; a month later the product shipped that capability, and while you were preparing for a sales review the model connected the two facts and flagged the customer as worth approaching again. In the second, the model works from authorised historical records to pull together what a predecessor knew about each client, how far each deal had progressed, what had been raised as feedback, so whoever takes over does not start from zero.

On the design side the output forms move a step further. The supported formats are Web, SVG, Three.js, Blender and PPT, of which animated PPT, animated posters and agentic 3D are new here. Animated PPT treats narrative pacing as part of the generation target: what each slide has to say, where it should pause, what deserves emphasis, and from that the transitions, the element entrance order and the way charts grow, all decided together rather than bolted on afterwards. The Blender path has an agent working inside Blender, producing a first version and then correcting against the actual render, fixing a floating roof or a misplaced pyramid and adding weather animation across several rounds, ending with a project file that can still be edited. The model also plugs into Figma and similar workflows.

The published comparisons include GPT-6 Luna, DeepSeek V4.1 Flash, GLM-5.3 Flash, MiniMax M3 and Claude 5.5 Haiku, over benchmarks that span in-house suites such as AutomationBench, WorkbuddyBench, TeamBench, DataAnalysis2.0 and RaccoonPPTBench v2 alongside public ones including BrowseComp, LongMemEval-S, BEAM 128K and PersonaMem-v2. SenseTime's conclusion is that it exceeds some frontier international models in agent collaboration, data analysis and presentation generation.

Key Features

• Long-term memory with proactive interaction: The model keeps not only the current turn but also your earlier goals and constraints. When related information changes it looks for what is affected and proposes adjustments, instead of waiting for you to remember and ask again.
• Work that continues across tasks and sessions: One job can run over several days without re-explaining the background when you switch sessions or entry points. SenseTime frames this as a task carried across days, which is what lets it treat office work as the main arena.
• Animated PPT that paces the narrative first: Narrative pacing is part of the generation target: what each slide conveys, where it should pause, what should be emphasised, and from that the transition types, the entrance order of elements and the way charts grow. In the delivered PPTX the motion and the page structure correspond to each other.
• Animated posters: Key visuals, text hierarchy and motion choreography are planned in one pass. How far the main visual moves and when text appears are tuned against each other, so the timeline serves the information hierarchy. Results work for social distribution and extend into a series.
• Editable 3D projects inside Blender: The agent produces a first version of the scene, then revises against the actual render. A floating roof or a misplaced pyramid can be raised item by item, weather animation can be added later, and after several rounds what comes out is a Blender project you can keep editing rather than a locked video.
• One model across many deliverables: Websites, SVG vector animation, Three.js web 3D, Blender scenes and PPT all come from the same model. The vector path outputs files and animation code that scale losslessly and can be edited again.
• Slots into existing design workflows: It connects to professional design tools such as Figma, so generated files land on the designer's existing canvas for further adjustment, linking generation, editing and iteration.

Use Cases

• Sales and account work, matching a product update against past opportunities and surfacing the customers worth revisiting.
• Handovers, turning client background scattered across chat, documents and email into readable handover notes.
• Content teams that need animated layouts, including launch decks, social posters and series extensions.
• First drafts of 3D scenes, letting the agent set structure and materials before refining in Blender.
• Design teams wanting a complete proposal before Figma refinement, cutting the prep time ahead of discussion.
• Complex work that runs across days, for instance a chain from research to draft to artwork.

Pros

• Long-term memory and proactive interaction land in concrete office scenarios rather than staying an abstract claim.
• Animated PPT treats narrative pacing as a generation target, fixing motion and page structure together.
• The Blender path delivers editable project files that can be refined further.
• One model covers websites, vectors, web 3D, Blender and presentations.
• Connects to existing design workflows such as Figma.
• The Token Plan is free during public beta, and both Studio and SenseNova Raccoon let you start immediately.
• The same release also open-sourced SenseNova Loop-9B-Preview, with the architecture exploration available to download.

Pricing

There are three official entry points: SenseNova Studio for online use (the Design experiment build), the SenseNova Token Plan, and SenseNova Raccoon. The Token Plan is entirely free during public beta at 0 CNY per month; the tier lists SenseNova 6.8 Flash Lite and SenseNova U1 Fast, allows up to 20 API keys per account, and paid tiers have not launched yet. On the Raccoon side the model is free until 1 December 2026. API documentation lives at platform.sensenova.cn/docs and the official skills repository at github.com/OpenSenseNova/SenseNova-Skills, both open to use.

Summary

Whether it is worth trying depends on what kind of work you have. If every request is a single question that ends with an answer, the difference between models shows up mainly in answer quality. If a job runs over several days and someone else picks it up partway through, memory and proactive interaction decide the experience, and that is what this release leans into.

Both office cases SenseTime showed sit in that direction. A customer declined because a capability was missing, the capability shipped a month later, and the model links the two while you prepare for a review. That comes from keeping context and actively looking at it, not from doubling the context window. The design changes extend the same logic: animated PPT thinks about narrative before motion, the Blender path looks at renders and revises across rounds, and what comes back is a project you can keep working on rather than a finished film.

Benchmark provenance is worth noting. Of the suites listed, AutomationBench, WorkbuddyBench, TeamBench, DataAnalysis2.0 and RaccoonPPTBench v2 come from SenseTime's own evaluation system, while the public set is BrowseComp, LongMemEval-S, BEAM 128K and PersonaMem-v2, and the official claim is exceeding frontier international models in some scenarios. The fastest check for your own work is running one real document through Studio.

Trying it costs little right now. The Token Plan is free during public beta, both Studio and Raccoon are open, and the Raccoon free window runs to 1 December.

Version History

• SenseNova 6.8 Flash ships, upgrading both office and design work (2026-11-02): SenseTime released the production version of SenseNova 6.8 Flash, optimised around complex real-world tasks in task execution and result delivery and supporting collaboration across tasks and sessions. On the office side long-term memory and proactive interaction let it identify affected items when information changes and suggest adjustments; on the design side it covers Web, SVG, Three.js, Blender and PPT, adding animated PPT, animated posters and agentic 3D design, and connects to Figma. Published evaluations span AutomationBench, WorkbuddyBench, TeamBench, DataAnalysis2.0, RaccoonPPTBench v2 plus BrowseComp, LongMemEval-S, BEAM 128K and PersonaMem-v2, with SenseTime claiming it exceeds some frontier international models in agent collaboration, data analysis and presentation generation. Available from launch through SenseNova Studio, the SenseNova Token Plan and SenseNova Raccoon, alongside the open-source SenseNova Loop-9B-Preview.