HomeAI models
Perceptron: Perceptron Mk1
•
Perceptron Mk1
•
Released 
May 2026


Perceptron Mk1

Video and embodied-reasoning vision-language model that returns timestamps, points, boxes and polygons from images and video at low cost.

Modality:
Text
Image
PDF
model ID
perceptron/perceptron-mk1
Start for free
arrow black icon
View API Docs
arrow white icon

Output Speed *

N/A
tok/s

Intelligence Index *

N/A
/ 100

Context Window *

32768
tokens

Input price

0.9
Anytoken

Output price

9
Anytoken
Perceptron Mk1: Video Understanding With Grounded Spatial Output Perceptron Mk1 is a closed-source vision-language model from Perceptron, released May 12, 2026 as the first model in the company's closed-source family. It accepts text, image and video input and returns text. On request it can also emit points, boxes, polygons and video clips. Reasoning is hybrid and can be switched off when it isn't needed. Mk1 fits video QA, event detection, spatial grounding and document OCR. Perceptron's Mk1.5 has since succeeded it. Access Perceptron Mk1 through the AnyAPI.ai API

Performance

Spatial Grounding and Video Reasoning, by Perceptron's Own Evaluations

Mk1 is built for perception tasks rather than general text work. Perceptron reports 72.4 on RefSpatialBench, against 2.2 for Claude Sonnet 4.5, and 88.5 on VSI-Bench, the highest score among the models it compared. These are vendor-published results, not independent tests. They suggest strong referring-expression and spatial grounding, which matters when downstream code consumes coordinates or timecodes. In production, Mk1 fits the perception stage of robotics and video pipelines, where its lower-cost tier supports high-volume analysis.

Technical Specifications

What the model supports

Mk1 takes text, image and video input and returns text, with optional structured annotations. The context window is 32K tokens, and video is analyzed at up to 2 FPS inside that window. Maximum output is a separate limit, listed at 8,192 tokens by third-party catalogs. The 32K window and 2 FPS ceiling set the practical clip length per request, so long footage needs segmentation. Function calling is not documented for Mk1.
Verified Specifications — 
Perceptron Mk1
*
Input modalities
Text
Image
PDF
output modalities
Text
Context window
32768
 tokens
Maximum output tokens
8192
Reasoning
Yes
Knowledge cutoff
May 2026
Pricing (standard)
0.9
 AnyTokens in
 / 
9
 AnyTokens out

Limitations & Trade-offs

Where Perceptron Mk1 falls short

1
32K context caps video length. Video runs at up to 2 FPS inside a 32K-token window shared with prompt text, and output is capped at 8,192 tokens per third-party listings. Long surveillance, broadcast or teleoperation footage must be split into segments, and events spanning segments need your own stitching logic. For long-form video where one request must cover extended footage, a model with a much larger context window is the safer choice.
2
Narrow evidence base beyond perception. Perceptron's published results cover video, spatial and image benchmarks, and no coding, agentic or general text benchmarks were found. Mk1 outputs text only and has no audio input. Function calling is not documented for it. Teams building general chat, code generation or tool-driven agents should use a generalist LLM and treat Mk1 as a specialist perception component.
3
Benchmarks are vendor-reported. The headline figures, such as 88.5 on VSI-Bench and 72.4 on RefSpatialBench, come from Perceptron's launch materials, and the comparisons used models current at that time. No independent speed or latency measurements were found. Treat the scores as directional and validate on your own footage, camera angles and annotation formats before committing to production.
4
Superseded by Mk1.5. Perceptron released Mk1.5 on September 25, 2026 with audio input, tracking, function calling, web search and reported 2-5x faster end-to-end requests. Mk1 remains available by explicit model ID, but new capabilities land on Mk1.5. Teams needing tracking across frames or audio-aware video analysis should evaluate Mk1.5 rather than Mk1.

Best-Fit Workloads

Where this model earns its place

01

Video QA and Event Clipping


Mk1 reasons across time and returns structured timecodes for specific moments in a video. Perceptron cites auto-clipping grasp attempts in teleoperation episodes and flagging restocking events in warehouses in a single prompt. It fits monitoring, sports-review and media-indexing pipelines built on short clips. The 32K-token window and 2 FPS ceiling mean long streams must be segmented first.

02

Robotics and Spatial Grounding


Mk1 is trained to output points, boxes and polygons as first-class results, so a downstream policy can use them directly. Perceptron-reported scores of 85.1 on EmbSpatialBench and 72.4 on RefSpatialBench support grasp pointing, multi-view understanding and success detection. It is a perception layer, not an action planner: pair it with control logic.

03

Document and Image Extraction


Perceptron says Mk1 converts complex documents to HTML, JSON or Markdown with layout and tables preserved, including multilingual text and handwritten annotations. JSON Schema structured outputs make it suitable for invoice, form and label parsing from photos. These are launch claims, so test on your own document mix. Text-only PDFs are better served by a text model.

04

High-Volume Visual Labeling


Perceptron positions Mk1 at the cost of lightweight models while matching frontier performance on its chosen benchmarks. That suits offline curation of teleoperation episodes, open-vocabulary detection and counting, and captioning at scale. Reasoning can be switched off for simple labeling to reduce latency. Expect to run quality sampling, since the benchmark claims are vendor-reported.

Pricing in anytokens via AnyAPI
Input
0.9
₳
Output
9
₳
Cache write
—
₳
Cache read
—
₳
Get API access
arrow black icon
Read Docs

on this page

PerformanceTechnical SpecificationsLimitations & Trade-offsBest-Fit WorkloadsIntegrationFAQ

Integration

Access Perceptron Mk1 via AnyAPI.ai

Access Perceptron Mk1 through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate Perceptron Mk1 without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

01

One API integration

Access Perceptron Mk1 and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

02

Easy model switching

Test Perceptron Mk1 against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

03

Flexible for production

Use Perceptron Mk1 from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

04

Multi-model applications

Use Perceptron Mk1 for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Start free
arrow black icon
Read Docs

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

Perceptron Mk1 has a 32K-token context window (32,768 tokens), which Perceptron documents as multimodal context. Video is analyzed at a dynamic frame rate of up to 2 FPS within that window. Maximum output is a separate limit, listed at 8,192 tokens by third-party catalogs. Long videos need to be split into segments.

Yes, Perceptron Mk1 accepts video natively, along with images and text. Documented video formats are MP4 and WebM; image formats are JPEG, PNG and WebP. Output is text, and it can include timestamped clips when requested. It does not accept audio input; that was added in Mk1.5.

Yes, Perceptron Mk1 is a hybrid reasoning model. Perceptron says reasoning can be turned off when it isn't needed. Turning it on suits multi-step temporal or spatial questions, at the cost of extra latency. On Perceptron's native API, the toggle is vision_config.enable_thinking.

Function calling is not documented for Perceptron Mk1, and OpenRouter's Mk1 endpoint does not accept tools. JSON Schema structured outputs are listed instead. Function calling is listed for the successor, Mk1.5. If your agent needs tool calls, pair Mk1 with a tool-capable model or evaluate Mk1.5.

No. Perceptron released Mk1.5 on September 25, 2026 as Mk1's successor. It adds audio input, tracking, function calling and web search, and Perceptron reports 2-5x faster end-to-end requests. Mk1 launched May 12, 2026 and can still be pinned by its model ID, which suits pipelines already validated on it.

* Benchmark data source: Artificial Analysis artificialanalysis.ai