SenseTime co-founder and chief scientist Lin Dahua has offered a deliberately conditional forecast: the “emergence moment” for multimodality could arrive within one or two years. KrASIA published its report of the interview on September 17, 2026, under the headline that a multimodal AI breakthrough could come within two years. This article is an editor-approved recency exception: the KrASIA report does not satisfy EastFrontier’s ordinary 24-hour eligibility rule. The forecast is Lin’s assessment in an interview, rather than a declared product deadline, benchmark result, or guarantee that a technical breakthrough will occur.
Lin Dahua’s Case for Vision-and-Language Models
The 36Kr Europe interview identifies Lin as SenseTime’s co-founder and chief scientist and places his prediction within a broader argument about where AI can be used. He said office work, healthcare, retail, robotics, and world models all require visual signals. In that framing, models that jointly handle vision and language matter because they could address tasks outside a text-only interface. “We believed very early that the joint modeling of vision and language is the foundation for artificial intelligence to enter real industries and the physical world,” Lin said in the interview.
Lin contrasted that ambition with coding agents’ current commercial momentum: coding has abundant process data, executable outputs, deterministic test feedback, and an established software-engineering market. He did not present coding as the definition of intelligence. “Strong coding capability is the result of improved intelligence level, but the core essence of intelligence is not the work form of coding, but to integrate complex information for long-range thinking, deduction, execution and correction,” he said. In SenseTime’s stated direction, visual perception and feedback become more important as systems address physical-world and industrial tasks. That rationale is Lin’s view, not a neutral industry consensus.
The “Alignment Tax” Behind Native Multimodal Training
Lin calls the need to add visual data without reducing language, mathematics, or reasoning an “alignment tax.” In the 36Kr interview, he said the competitive penalty of an additional experiment could be lost time: “For model companies, the problem may not be that they cannot afford the cost, but that one extra round of experiment will make the model one month slower than competitors.” The remark describes a development-speed concern, not a disclosed cost figure.
SenseTime’s response, according to the interview transcript, is to begin iterations with lightweight models containing 8 billion parameters because the unified approach to understanding and generation is not yet settled. Lin described an experimental route rather than an already resolved recipe for native multimodality. Separately, the emergence moment remains a forecast rather than an established outcome. The earlier EastFrontier report, SenseTime Releases SenseNova-U1, an Open-Source Image Model That Reasons in Pictures, provides related coverage of the U1 model family.
SenseNova U1’s Unified Representation Design
In an April 28 SenseTime research announcement, the company described SenseNova U1 as a native unified multimodal model series that combines multimodal understanding, reasoning, and generation within a monolithic model framework. SenseTime said its NEO-Unify architecture removes the visual encoder and variational autoencoder, or VAE, and instead uses a unified representation space for language and images. These are first-party architectural claims from SenseTime, not independently verified evidence of comparative model performance.
The 36Kr interview and KrASIA report describe the same U-series direction in product terms. They say U1.5 Lite extends the approach to continuous editing, high-resolution generation, and complex instruction control. SenseTime’s own U1 announcement says the U1 Lite release includes 8B and A3B mixture-of-transformers configurations and is open source. The available sources do not support claims that the series leads named rivals or achieves a particular benchmark position, so the relevant fact is the company’s stated design and release description rather than an unverified performance ranking.
SenseNova 6.8, U-Series Products, and Agent Strategy
KrASIA characterizes SenseTime’s strategy as an expansion from its computer-vision origins into native multimodal AI and spatial intelligence. Its report describes SenseNova 6.8 as a set of multimodal agent models intended for extended-task planning, execution, and multi-agent coordination. It separates that line from the U series, which the report says focuses on native visual understanding, generation, and editing, with extensions into video and 3D. KrASIA also names Seko, an AI agent linking storytelling, storyboarding, and video generation, and Office Raccoon, an office-work and data-analysis agent. Those are KrASIA’s reported descriptions of the company’s product strategy, not independently audited results.
Lin put a longer-range objective on the U series in unusually direct terms: “The U series has a longer-term goal, which is spatial intelligence and physical intelligence.” That statement connects the interview’s premise—joint vision-and-language modeling—with SenseTime’s stated product direction. It does not establish that SenseTime has reached spatial or physical intelligence, and it does not turn the two-year forecast into a delivery commitment. Readers seeking related context on the company’s positioning can see SenseTime Bets on Cost Efficiency and Overseas Markets to Stay Relevant in China’s AI Race.
Company-Reported First-Half 2026 Financial Context
SenseTime’s August 26 interim-results announcement reported first-half 2026 revenue of RMB 2.91 billion, a 23.4% year-over-year increase. The company reported RMB 2.33 billion in generative-AI revenue, nearly 80% of total revenue, and RMB 620 million in IFRS net profit, which it described as its first IFRS profit since listing. It also listed NEO-Unify, SenseNova U1, U1.5, and U1.5 Lite among its 2026 model releases.
Those figures provide commercial context, but they do not demonstrate that multimodal AI alone caused the reported profit. SenseTime attributes the result to a broader AI and commercialization framework that includes native multimodal models, agents, and infrastructure. The financial disclosure is company-reported interim information, not an independent audit in the supplied record. For earlier coverage focused on those reported results, see SenseTime Turns Its AI Bet Into a First-Half Profit.
A Two-Year Forecast, Not a Multimodal Milestone
The significance of Lin’s interview is therefore narrower than a launch announcement. It records how SenseTime’s chief scientist links native multimodal modeling, the alignment tax, small-model iteration, U-series architecture, and a longer-term spatial and physical intelligence goal. It also records his expectation that an emergence moment could arrive within one or two years. KrASIA’s September 17 report and the August 28 36Kr interview support attribution of that forecast to Lin; neither source establishes that a breakthrough has already happened or will definitely happen by a specified date. The company’s U1 description and financial release add stated technical and commercial context, while leaving that central prediction exactly where it belongs: with the speaker who made it.