Perceptron Mk1 — Initial public release¶

Development line: project:perceptron-mk1 · thread initial-public-release
Last event: 2026-05-14 · 1 dated since 2026-05-14 · Researched: 2026-09-06 · confidence: medium

What it is¶

Perceptron Mk1 is a closed vision-language model for teams adding image and video analysis to APIs, not a video generator.

• Multimodal input: accepts text, images, and video.
• Structured output: returns text with spatial or temporal annotations.
• Visual tasks: handles Q&A, OCR, object detection, captioning, video-event clipping, and in-context visual examples.
• Configuration: sets vision_config for reasoning and typed grounding.

Development line¶

• 2026-05-14 — Perceptron Mk1 public introduction. On 2026-05-14, Perceptron Mk1 appeared on a public introduction page. The post links to a demo, documentation, and an OpenRouter model page, marking a material public development milestone. Those links alone do not establish model capabilities, technical design, or access terms.

What changed¶

2026-05-14 — Perceptron published the Mk1 announcement; official release history dates the actual launch to 2026-05-12, adding image/video input, 32K context, reasoning, model specifications, and video guides. 2026-06-09 — A dedicated /v1/detect endpoint added direct grounded detection from categories, exemplars, or exhaustive requests. 2026-06-15 — A Files API added reusable uploaded image and video references across requests. 2026-06-22 — Chat completions gained timestamped inline video_frames as an alternative to a video URL.

How to use this¶

As of 2026-05-14, start with the public documentation and demo, then verify current OpenRouter availability and terms before adoption.

1. Open the demo, sign in if prompted, and test a concrete prompt on a sample asset before integrating. — https://www.perceptron.inc/demo
2. Create a Perceptron API key, select perceptron-mk1, and use the Python SDK, cURL, or an OpenAI-compatible client. — https://docs.perceptron.inc/
3. Send text plus a JPEG, PNG, WebP, MP4, or WebM asset to the chat-completions API. — https://docs.perceptron.inc/
4. Choose the result contract upfront: enable thinking for reasoning tasks, or request point, box, polygon, or clip annotations for grounded output. — https://docs.perceptron.inc/perceptron-mk1/models/perceptron-mk1
5. For assets reused across requests, upload them once and reference the file ID; each file is capped at 128 MiB. — https://docs.perceptron.inc/perceptron-mk1/guides/files
6. As a routed alternative, select perceptron/perceptron-mk1 through OpenRouter and check the current model card before deployment. — https://openrouter.ai/perceptron/perceptron-mk1

Best practices¶

• Use observable prompts and typed output requests: named classes for detection and an exact event predicate for clipping. — https://docs.perceptron.inc/perceptron-mk1/best-practices/prompting-reference
• Enable thinking for text Q&A, OCR, captioning, and clips; leave it off for point, box, and polygon localization. — https://docs.perceptron.inc/perceptron-mk1/models/perceptron-mk1
• For video clips, ask for the exact moment or range; treat a returned single timestamp as an approximate instant rather than a zero-length interval. — https://docs.perceptron.inc/perceptron-mk1/capabilities/video-clipping
• Keep interactive timeouts and max_tokens tight; resize media client-side and honor Retry-After with jitter after a 429. — https://docs.perceptron.inc/perceptron-mk1/guides/scaling
• Store keys in environment variables or a secret manager, strip PII from persisted prompts and logs, and protect annotated examples as proprietary imagery. — https://docs.perceptron.inc/perceptron-mk1/best-practices/security

Superseded by this¶

• 2026-06-22 — None documented: the public changelog describes the June capabilities as additions without deprecating Mk1, chat completions, or video_url input.

Still unknown¶

• The supplied launch article at https://www.perceptron.inc/blog/introducing-perceptron-mk1 returned 403 during this check, so its text and publication timestamp were not used as primary proof.
• The 2026-05-14 date may reflect when the post appeared rather than the actual product launch; unavailable private text prevents confirmation.
• No later Mk1 model revision appears in the public changelog after 2026-06-22, though undocumented service changes remain possible.
• First-party documentation says 32K context while OpenRouter displays 33K; OpenRouter also specifies 32,768 tokens, so this looks like rounding rather than a material conflict.
• No first-party Simplified-Chinese release was found. The cited Chinese roundup corroborates launch claims without independently validating vendor performance claims.

Sources¶

Agent brief¶

• Subject: project:perceptron-mk1, thread initial-public-release, 1 dated events 2026-05-14 → 2026-05-14.
• Practical note: As of 2026-05-14, test the public documentation and demo first, then verify current OpenRouter availability and terms before adoption.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.