Perceptron Mk1: A Video Model That Answers in Coordinates

Most vision-language models are judged on how well they describe a picture. That is the wrong test for a robot, a factory camera or a sports clipping pipeline. Those systems do not want a paragraph about the scene. They want a coordinate, a bounding box or a timestamp they can act on without a human reading anything.

Perceptron Mk1 is built around that second job. Perceptron, a company that describes itself as "the model layer for physical AI", released it on 12 May 2026 as the first model in what it calls a new family of closed-source models. It is worth understanding even now that Mk1.5 exists, because the design choices are what decide whether either model fits your pipeline.

What Mk1 is, and what Perceptron has not said

Mk1 is a vision-language model that takes images, video and multiple camera streams alongside a text prompt. Perceptron calls it a hybrid reasoning model, meaning reasoning can be switched off when a task does not need it. The vendor states 32K tokens of multimodal context, and OpenRouter lists a maximum output of 8,192 tokens.

Perceptron has not published a parameter count, an architecture description or a licence for the weights. Mk1 is closed: you reach it through the Perceptron Platform and SDK under the model id perceptron-mk1, or through a router such as OpenRouter. The company's earlier Isaac models were released openly, and Perceptron says Mk1 surpasses that series "across the board". If you need weights you can host yourself, Mk1 is not the answer; Perceptron's site says commercial licences to its weights are a sales conversation.

Coordinates, not captions

The interesting part of Mk1 is its output vocabulary. Perceptron says it was trained with structured spatial outputs alongside text: points, boxes, polygons, tracks and clips. That changes what you can build:

• Pointing. Ask where the grasp point on an object is and get a coordinate a controller can use, including for tasks like hand pose estimation.
• Counting. Perceptron claims it counts reliably "into the hundreds", the range where you should check any vision model's answer against a hand count.
• Temporal grounding. Return timecodes for specific moments in long footage, such as auto-clipping grasp attempts in teleoperation recordings.
• Instrument reading and OCR. Analog clocks, gauge needles and messy text, which matters for industrial inspection.
• Document extraction. Conversion to HTML, JSON or Markdown with layout preserved.

Through an OpenAI-compatible route the response arrives as text, so the geometry is serialised inside it. Before you write a parser, send a handful of requests and look at exactly how points and boxes come back on your route, then request structured outputs where the endpoint supports them. Best model for structured output covers why constraining the format beats hoping for it.

The 32K budget is the real constraint

Perceptron samples video at a dynamic frame rate of up to 2 frames per second, and every frame costs tokens. A 32K context sounds generous for text and is tight for footage. A long clip will not fit whole, and multi-camera input divides the same budget between streams.

The practical consequence is that Mk1 is a model for clips, not archives. If your job is "find every forklift near-miss in eight hours of warehouse video", the architecture around the model matters more than the model: cut the footage into segments, run them independently, and merge the timecodes afterwards. Budget for that orchestration before you compare prices.

Speaking of price, Perceptron lists Mk1 at $0.15 per million input tokens and $1.50 per million output, and OpenRouter shows the same figures at the time of writing. Input is the side that grows with video, and it is the cheap side. For grounding tasks the output is short, so per-clip cost is usually dominated by frames rather than answers.

What Mk1 does not do

Three gaps are worth knowing before you commit.

• No tool calling on OpenRouter. The OpenRouter listing for Mk1 does not accept tools or tool_choice. Mk1 perceives; it does not drive an agent loop by itself. Mk1.5 added tool use.
• No audio. Inputs are text, image and video. If the sound of a machine or a spoken instruction matters, Mk1 cannot hear it.
• Benchmarks in charts only. Perceptron's launch post claims Mk1 matches frontier models from Google, Alibaba, OpenAI and Anthropic on video and embodied reasoning at lower cost, but presents the results as charts and notes that some frontier-model points are noisy single-point estimates. Treat it as a claim to verify, not a ranking.

Mk1 or Mk1.5

Both models are listed at the same price, so for new work the default should be Mk1.5: Perceptron reports it is two to five times faster end to end and adds audio, tracking and tool calling. The read on that model is in the Perceptron Mk1.5 guide.

Mk1 still makes sense in one situation: you already have a pipeline tuned against its output format and thresholds, and a model change means re-validating every downstream rule. OpenRouter lists Mk1 under a dated canonical slug, perceptron-mk1-20260512, which is the point of a snapshot: behaviour you validated stays put. Migrate deliberately, with a comparison run, rather than because a newer number appeared.

How to evaluate it on your own footage

Generic vision benchmarks say little about whether Mk1 can read your gauges or count your parts. A short evaluation does:

1. Collect 30 clips or images from your real cameras, including bad lighting, occlusion and motion blur.
2. Label the ground truth yourself: the count, the box, the timecode.
3. Measure error in units that matter downstream, such as pixels off target or seconds off the event, not a pass or fail.
4. Run the same set through one general multimodal model as a baseline. Vision-language models compared lists what to test.

If Mk1's coordinates land inside your tolerance and the general model's do not, the specialist earns its place. If both miss, the problem is probably your footage or your prompt, and no model swap will fix it.