Perceptron Launches Mk1 Vision-Language Model with Video Reasoning at $0.15/$1.50 per 1M Tokens
Perceptron has released Perceptron Mk1, a vision-language model designed for video understanding and embodied reasoning tasks. The model accepts image and video inputs with 33K context window, priced at $0.15 per 1M input tokens and $1.50 per 1M output tokens, and supports structured spatial annotations on demand.
Perceptron Mk1 — Quick Specs
Perceptron Launches Mk1 Vision-Language Model with Video Reasoning
Perceptron has released Perceptron Mk1 (Mark One), a multimodal vision-language model built for video understanding and embodied reasoning tasks. The model processes image and video inputs paired with natural language queries, returning either structured annotations or natural language responses.
Pricing and Context
Perceptron Mk1 is priced at $0.15 per 1M input tokens and $1.50 per 1M output tokens, with a 33K token context window. The model is available through OpenRouter's API routing service.
Core Capabilities
According to Perceptron, Mk1 excels at multiple video understanding tasks including video question answering, summarization, and event detection. For image inputs, the model handles:
- Point-by-example grounding from multimodal prompts
- OCR and document parsing on real-world inputs
- Open vocabulary object detection and counting
- Hand pose estimation
Structured Annotation System
The model's distinctive feature is its optional structured annotation output. By default, Mk1 returns natural language text only. Users can request spatial localization through the annotation_format parameter:
- "point" for point annotations on images
- "box" for bounding boxes
- "polygon" for polygon masks
- "clip" for temporal segments (start/end timestamps) in video
Annotations are emitted inline with text only when explicitly requested.
Optional Reasoning Mode
Mk1 includes an optional reasoning mode that can be enabled per request. This trades increased latency for deeper analysis on complex tasks, allowing the model to show step-by-step thinking processes. OpenRouter provides access to the reasoning_details array in API responses.
What This Means
Perceptron Mk1 enters a crowded multimodal model market with a focus on structured output formats and video understanding. The $1.50 per 1M output tokens places it in the premium tier—comparable to GPT-4 Vision pricing. The optional reasoning mode and granular annotation controls suggest the model targets developers building computer vision pipelines and video analysis applications rather than general-purpose chat interfaces. The company has not disclosed benchmark scores or parameter count, making direct performance comparisons difficult.
Related Articles
Meta Releases Muse Glimmer 30B, an On-Device Agentic Model with Built-In Perception Encoder
Meta Superintelligence Lab has released Muse Glimmer, a 29.6-billion-parameter multimodal model distilled from Muse Spark for autonomous agentic tasks that run entirely on consumer hardware. The Apache 2.0-licensed model ships with a dedicated perception encoder, 131K+ token context, and speculative decoding for local speedups up to 3.1x.
Ollama Adds Support for Meta's 30B "Muse Glimmer" Model, First Release from Meta Superintelligence Labs
Ollama v0.32.7 adds support for Muse Glimmer, a 30B-parameter multimodal model that Meta describes as the first release from its new Meta Superintelligence Labs division. Initial support runs through Ollama's MLX engine on Apple Silicon, with NVIDIA and AMD support promised in the coming days.
Meta Releases Muse Glimmer, a 30B Multimodal Model Distilled for Local Agentic Use
Meta has released Muse Glimmer, a 30B parameter multimodal model distilled from its larger Muse model, designed for local, privacy-aware agentic applications. The model ships under the Apache 2.0 license as part of Hugging Face Transformers v5.15.0, which also adds support for IBM's GraniteSWA/GraniteMoeSWA and the A.X-K1/K2 models.
Mistral AI Releases Shieldstral-1.0-3B, a 3B-Parameter Policy-Adaptive Safety Classifier
Mistral AI has released Shieldstral-1.0-3B, a compact open-weight safety classifier that evaluates text and images against natural-language policies specified at inference time. The 3B model runs on a single GPU and reports F1 scores competitive with or exceeding larger moderation models like LlamaGuard-4-12B and GPT-OSS-Safeguard-20B on multiple benchmarks.
Comments
Loading...