MediaTek Puts Camera Coaching on the Dimensity Chip

MediaTek and vivo put a vision-language model in the Dimensity 9500 camera viewfinder, running on the phone in milliseconds. They showed that Camera AI Agent at Mobile World Congress on March 4, 2026. It reads the scene before the shutter, then offers composition help and vivo Signature Effects with no network in the loop.

Google already ships a similar guide on Pixel 10 phones. That guide still asks for a data link. The first Dimensity 9500 phones that can host MediaTek’s local model, the vivo X300 and X300 Pro, had already gone on sale the previous October.

A Coach That Reads the Scene in Milliseconds

MediaTek said it brought the first millisecond-level real-time vision-language model, or VLM, to the Dimensity 9500. The model sits on the handset. It looks at whatever the sensor is seeing, names the scene, and answers fast enough to keep up with a live preview. Nothing in that loop is described as a cloud round trip.

That is a different job from the gallery editors most flagships already ship. Those tools rewrite a photo after you take it. This agent is meant to speak while you are still framing, so the advice can change the shot instead of the file.

WHAT THE AGENT DOES IN THE VIEWFINDER

• Scene reading: A local VLM studies the live preview and classifies what is in front of the lens.
• Shooting guide: It suggests framing, light, and style the way a coach would talk a novice through a setup.
• Auto composition: It can propose a tighter crop or a shift in the frame before you tap the shutter.
• Signature Effects: It offers a selection of vivo graphic styles as part of capture, not as a filter you paint on later.
• Offline path: Inference stays on the Dimensity 9500, so a dead radio does not kill the guide.

The pitch at Barcelona was that this stack can spread with the chip. Any brand that ships a Dimensity 9500 could, in theory, hang its own camera app on the same VLM. vivo is only the first skin.

Pixel’s Camera Coach Still Needs a Network

Google built the closest public cousin on the Pixel 10 series in 2025 and called it Camera Coach. It also talks before the shutter. It also claims to teach framing, light, zoom, and mode choice. The plumbing is not the same.

Google’s own how-to is blunt about the network. Molly McHugh-Johnson, writing on the Pixel blog, told owners to get online first because the feature uses cloud-based Gemini models for the most relevant tips.

First, make sure you’re connected to data or Wi-Fi because Camera Coach uses some cloud-based Gemini models to give you the best, most relevant suggestions.

Molly McHugh-Johnson, Pixel, Google

Pixel Camera Help matches that limit. Camera Coach is off until you turn it on, and the guide needs an internet connection. If the radio is down, the coach is down. The MediaTek demo was sold as the opposite of that: a model light enough, and an NPU fast enough, to keep the conversation inside the phone.

TWO PRE-SHUTTER COACHES

Pixel still has the deeper photography software stack, from Night Sight to the later editing tools. What it does not have is a chip it can sell to OPPO or Xiaomi. MediaTek does. Once a VLM is a platform block, coaching stops being a Pixel party trick and becomes another ISP feature brands can skin.

What the Dimensity 9500 NPU Runs on the Phone

The agent is not a magic camera app. It is a model that only works if the NPU can score a live viewfinder without melting the battery or dropping frames. MediaTek built the Dimensity 9500 for that class of job, with a dual NPU setup it calls Super Performance plus Super Efficient, the second block using compute-in-memory for always-on work.

The headline unit is the NPU 990 with Gen-AI Engine 2.0. MediaTek says it is twice as powerful as the previous Dimensity flagship’s NPU, cuts peak power use by 56%, and more than doubles token generation speed. It also runs 4K text-to-image on the device. Light models on the CIM block draw 42% less power than a typical NPU, which is the kind of budget you need if a VLM is going to watch the preview all the time.

NPU 990 AND IMAGIQ 1190

• Token speed: Over 2x faster generation than the prior Dimensity flagship NPU, per MediaTek Labs.
• Peak power: Up to 56% lower at peak AI load, which is why a live VLM is even plausible.
• Focus tracking: The Super Efficient NPU feeds 30 FPS focus tracking, against a 6-10 FPS industry habit.
• Context window: On-device long-text capacity moves from 32K tokens on the Dimensity 9400 to 128K, a 4x jump, or about 10 hours of recording context against 2.5 hours.

The imaging side is the Imagiq 1190 ISP. It does RAW-domain processing before compression, 200MP capture, 4K120 recording with dual-track stabilization, and 4K60 cinematic portrait video. Those are shipped camera features. The MWC agent is the language model sitting on top of that pipe, reading the same preview the ISP is already chewing.

Around the NPU, the chip is a third-generation all-big-core design on TSMC’s N3P node: one C1-Ultra at 4.21 GHz, three C1-Premium cores at 3.5 GHz, and four C1-Pro cores at 2.7 GHz. MediaTek lists 32% faster single-core and 17% faster multi-core results against the Dimensity 9400, with 37% lower multi-core peak power and 55% better Ultra-core peak efficiency. The GPU is a 12-core G1-Ultra. That is the computer. The agent is software that finally has a place to live.

Signature Effects Land Before the Shutter

vivo’s cut of the demo is not a neutral photography teacher. After the VLM names the scene, the agent offers Signature Effects, a house set of graphic looks designed to land as you shoot. The style is part of the capture, not a slide you drag later in a gallery.

That is the OEM lock hiding inside a platform feature. MediaTek can sell the same millisecond VLM to whoever buys the chip. The look of the photo still belongs to vivo if those effects are the default path from preview to JPEG. A rival brand would need its own pack, or users would keep taking pictures that look like vivo pictures on a different badge.

The coach half is closer to Pixel’s idea: composition, light, and style tips for people who do not already think in thirds and catchlights. The difference is timing and ownership. Pixel’s coach tries to teach you, then leaves the pixels alone. vivo’s agent can also paint the brand on the file before the file exists.

People who follow these phones already treat the look as the product. Launch talk around the X300 series stayed on the 200MP ZEISS modules, the V3+ block, and 4K60 portrait video, not on a named viewfinder model. The agent has to beat that habit if it wants to be more than a keynote slide.

Hardware the X300 Series Already Put on Sale

The awkward date in this story is not March 2026. It is October 2025. vivo’s global launch release on October 30, 2025, already put the X300 and X300 Pro on the record with the Dimensity 9500, OriginOS 6 on Android 16, and a dual-chip imaging path that pairs the SoC with vivo’s V3+ block. The Camera AI Agent arrived months later as software meant for phones that were no longer vapor.

FROM CHIP TO MWC DEMO

1. October 2025: vivo puts the X300 and X300 Pro on sale in China, then announces the global pair on October 30 with the Dimensity 9500 and V3+.
2. November 4, 2025: MediaTek publishes its joint-design note on the series, citing 32 percent faster single-core performance against the Dimensity 9400 and a partnership that began with the vivo X80 and Dimensity 9000 nearly four years earlier.
3. March 4, 2026: MediaTek and vivo show the Camera AI Agent in Barcelona and say it is coming soon to Dimensity 9500 vivo phones in China, with more technical detail promised in the following weeks.

The hardware the agent would ride is not a mystery. The compact X300 uses a 6.31-inch panel, weighs 190 g, is 7.95 mm thick, and carries a 6040 mAh typical cell (5925 mAh rated). The X300 Pro uses a 6.78-inch panel, is 7.99 mm thick, and carries a 6510 mAh typical cell (6375 mAh rated). Both list 90W wired charging and 40W wireless. vivo rates the Pro’s main and telephoto stabilization at CIPA 5.5 and the X300’s at CIPA 4.5. The Pro’s main camera is a 50MP Sony LYT-828 with a 1/1.28-inch sensor; both phones use 200MP ZEISS Ultra-Sensing HPB modules, the Pro as telephoto and the X300 as the main. Software support is listed at 5 years of OS upgrades and 7 years of security.

vivo already shipped an AI Image Studio on that global launch: AI One-Shot Multi-Crop, AI Landscape Master, and Live Photo Passerby Erase. Those are after-the-fact or parallel helpers. They are not the millisecond VLM in the viewfinder. The chip was jointly defined by vivo and MediaTek, and the same release called out the first 4K 60 FPS cinematic portrait video as the dual-chip party trick. The MWC agent is the next software layer on silicon that had already done a world tour.

OPPO’s Omni Model on the Same Chip

The same MediaTek keynote in Barcelona had another partner on stage. OPPO showed Omni, described as an on-device full-modal model that takes voice, video, and text and answers without a cloud hop. The Find X9 series, which also uses the Dimensity 9500, was the handset in the booth. OPPO said those on-device features would move through a ColorOS 16 update.

Omni is not vivo’s camera agent. It is the other half of MediaTek’s pitch: sell the NPU 990 as a place where brands run their own agents, then let each brand pick a job. OPPO took scene understanding and live Q&A. vivo took the viewfinder. A third brand could take either, or neither, and still ship the same SoC.

That is why the Pixel comparison is not a feature checklist. Google can keep Camera Coach as a Gemini product on Tensor phones. It cannot drop that coach onto an OPPO Find X9 or a vivo X300. MediaTek can. The loser, if the software actually ships, is not a camera sensor. It is the idea that pre-shutter AI advice is a Google exclusive.

China First, With Other Makers Still Waiting

At the March demo, MediaTek and vivo said the agent would arrive soon on vivo’s Dimensity 9500 phones in China, starting with the X300 series. They did not lock a week. They did not name a global build. They did not name a second OEM. More technical detail was supposed to follow in the weeks after the show.

By September 2026, six months after that Barcelona showing, the phones’ camera hardware is thoroughly documented and on shelves. A named, public rollout of the viewfinder VLM is not. That does not prove China never saw a quiet toggle. It does mean the platform story MediaTek told, a coach any Dimensity 9500 maker could copy, is still a slide until another brand ships its own skin.

WHAT WE KNOW

• The demo: A local millisecond VLM on Dimensity 9500, shown with vivo on March 4, 2026.
• The first phones: X300 and X300 Pro, already on sale from October 2025, listed as the debut vehicles in China.
• The contrast: Pixel Camera Coach still depends on a network and cloud Gemini models.

WHAT IS UNCONFIRMED

• Ship date: No later official note in the same channels pinned a week the agent went live, in China or abroad.
• Other markets: No confirmation the feature leaves China or lands on global OriginOS builds.
• Other OEMs: No second Dimensity 9500 brand has shown its own viewfinder agent, even though OPPO already runs Omni on the same silicon.

The chip can host the coach without phoning home. vivo can wrap it in Signature Effects so the first photos look like vivo photos. Until another maker actually ships that loop, Pixel still has the only coaching feature you can walk into a shop and tap, and it still needs Wi-Fi to speak.