SenseTime combines multimodal models with office and visual AI tools

SenseTime is an AI software company whose current offer includes multimodal models, workplace applications, computer vision and computing infrastructure. Its SenseNova model family is particularly relevant to teams turning documents and data into presentations or visual explanations. The key distinction is between a model call, a skill that operates tools and a finished application that manages the work.

01 / ProductSenseNova sits within a broader AI company

The company overview describes generative AI, Vision AI and SenseCore infrastructure as related business areas. This matters because a buyer investigating a presentation assistant is not automatically buying the compute platform or a computer-vision deployment. A shared corporate identity does not imply one contract, price or implementation process for every product.

The current product portfolio separates SenseNova foundation models and applications, SenseCore infrastructure and SenseFoundry industry products. Within SenseNova, the current FAQ names 6.8 Flash Lite, U1 Pro and U1.5 Lite, with different emphases. The model family should not be reduced to an old SenseChat description; the company says that product is no longer prominently listed but has not announced a formal discontinuation.

The SenseNova-Skills repository supplies modular office workflows for compatible agent runtimes. These cover presentations, spreadsheet analysis, research and visual output. A skill is application logic and instructions around tools. It is not a guarantee that the underlying language model alone can open a workbook, create an editable deck or preserve every number correctly.

02 / AudienceA useful candidate for analysts who must explain their work

A concrete audience is an operations analyst preparing a monthly service-performance presentation. The input may include an approved workbook, definitions of each measure and notes explaining unusual events. The output needs an understandable narrative, accurate charts and visible qualifications. An attractive slide is useful only if someone can trace its claims back to the accepted data.

The broader ecosystem may suit a team that wants reusable office workflows inside an existing agent environment, or one evaluating the packaged Raccoon experience. Those are different adoption paths. The former requires someone to own dependencies and tool permissions; the latter requires an assessment of the application’s workspace and commercial terms. Neither can be judged solely from a model demo.

The Gamma blueprint is a useful comparison when the primary need is presentation authoring and layout. The Genspark blueprint provides an adjacent approach to research and generated work products. Evaluate how each route preserves evidence and enables editing after generation, rather than comparing only the first exported image.

03 / WorkflowA proposed workflow keeps the data pack ahead of the deck

This proposed workflow starts with a frozen, approved workbook. Give each metric a definition, reporting period and unit. Separate actual results from forecasts and targets. Include a short list of permitted conclusions and known limitations, such as a service team that changed its reporting method halfway through the month. This is a designed evaluation, not a reported customer implementation.

First have the analysis stage produce a compact evidence table: measure, value, comparison period and source cell or sheet. Recalculate totals with ordinary spreadsheet or code tools. A model should explain the calculation and flag gaps; it should not become the only calculator for figures that will be presented to management. Keep the accepted table as the reference for every later format.

Then request a narrative outline. For example, one slide might distinguish a rise in overall requests from a rise in the proportion handled late. Those are different business statements even when both mention workload. Have the reviewer accept the outline before asking for layouts, so visual production does not harden a mistaken interpretation into a polished story.

The skills repository describes a presentation flow that collects a task pack and information pack before creating the deck. Use that separation to retain audience, purpose and source material as explicit inputs. Limit local file access to the evaluation folder. A workflow capable of reading and writing office files needs a narrower boundary than a simple chat box.

The U1.5 Lite integration announcement documents image generation and editing through image endpoints. It specifically warns that this image model cannot be configured as a chat model in coding tools. If the deck needs a visual explanation, call the appropriate image tool from the workflow rather than sending an image-model identifier to a text conversation endpoint.

Use generated imagery for explanatory scenes or conceptual diagrams whose details can be checked. Render exact financial or operational values from the approved evidence table. When a generated infographic contains text or counts, compare each item with the table before accepting it. The model’s ability to generate an attractive chart does not establish that its axes or numbers are correct.

Finish with a deck that the analyst can revise and an evidence packet that survives export. Measure incorrect values, unsupported conclusions, time spent repairing layouts and whether a second reviewer can locate the source of each important claim. Retain a manual route for the final presentation so a change in a beta quota cannot prevent the reporting deadline.

04 / PricingA free beta is a testing route, not a fixed operating price

The older FAQ describes 1,500 calls per model every five hours as of 27 August 2026. The directly retrieved current Token Plan page instead shows 60,000 points per five hours. These are different units, so this review does not convert one into the other or promise that the older call allowance remains available. The model-specific rules in the account are consequential.

A beta price does not establish the future cost of a recurring report. Count model calls, image revisions, analysis retries and the runtime used by the skills. Even when a model allowance is temporarily free, the team still supplies integration and review effort. Estimate the cost of the accepted deck when a paid plan becomes available, not merely the cost of one initial request.

The skills documentation identifies distinct international and mainland China endpoints and says the documentation, key, base URL and model name must all match the region. Do not move a Chinese quota assumption into an international budget without confirmation. Regional routing also belongs in the decision about which approved source files may enter the workflow.

05 / DistinctionsThe office-tool layer is a meaningful part of the offer

SenseTime’s combination of multimodal models and reusable office skills offers a concrete way to study the gap between answering a question and delivering an artifact. The skills can be inspected as workflow components, which helps a team understand where files are read, intermediate results are created and another model call is made. Inspection still needs to be followed by testing in the chosen environment.

The SenseNova-U1 repository describes unified visual understanding and generation, including a newer U1.5 release. This is relevant to iterative visual work, where the system must understand a reference and then revise an output. Open weights and a hosted API remain separate routes; availability of a repository does not mean its exact checkpoint is included in a subscription.

The strongest editorial distinction is the emphasis on completed office deliverables. That is a promising evaluation target because it can be judged directly: can another person edit the deck, reproduce the calculations and understand the argument? Vendor claims about token savings or production readiness do not answer those questions for your reporting process.

06 / QuestionsDense text, editing drift and region-specific access need testing

The U1 repository openly lists remaining problems with dense text, constrained layouts, exact counts and multi-step editing drift. For a business presentation, these are not cosmetic edge cases. A changed percentage or an extra box in a process diagram can alter the meaning. Review the actual deliverable rather than accepting the image because the overall composition looks plausible.

The U1.5 Lite announcement requires account registration and verification before key creation. This public review did not create accounts or confirm entitlement for a particular organization. It also did not validate Raccoon’s claimed security properties or a private deployment. Those claims require the corresponding product terms and a configuration-specific assessment.

Which parts of the deck remain editable after export? Can the chosen skill recover from a failed generation without re-running the analysis? Can the workflow retain the approved data table while replacing only a layout? These are useful pilot questions because they determine whether an analyst can correct a small mistake without regenerating the whole report.

07 / DecisionJudge the delivered presentation and its evidence together

SenseTime is worth considering when a team wants to combine analysis, narrative and visual output in a repeatable office workflow. Start with a modest reporting task and retain an independent reference for every important number. Choose the application or API route first, verify its regional access and quota, and evaluate the complete artifact before committing to recurring production.