Perceptron Mk1.5: a video model that outputs boxes, not just text

It returns points, polygons and timestamped video tracks your code can act on for $0.15/M — and Isaac 0.1/0.2 retire the same day, now 404ing.

Nowline SEP 27 12:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Geometry out, not captions

    Mk1.5 answers with machine-parseable geometry — points, boxes, polygons, `<track>` elements and video clips — returned inline via an `annotation_format` parameter, so your downstream code gets coordinates instead of prose to re-parse. It takes text, image, video and audio in, and Perceptron says one model drives embodiments from drones and quadrupeds to smart glasses and phones without platform-specific retraining.

  • It out-tracks Gemini, for pennies

    On Perceptron's own numbers, Mk1.5 hits 0.9433 hand-box localization vs Gemini 3.1 Pro's 0.4467, leads 3 of 4 video-tracking benchmarks (0.647 F1 on Molmo2-Track), and scores 74.7% on the DailyOmni audio-visual test. It runs $0.15 per million input tokens and $1.50 output, with cached input at $0.0375/M — and the Mk1 line was already reported 80-90% cheaper than OpenAI, Google and Anthropic on video.

  • Isaac is gone — unknown IDs now 404

    Perceptron retired the Isaac 0.1 and 0.2 model IDs the same day Mk1.5 shipped: they no longer resolve through the API or the hosted MCP server and now return HTTP 404. The same change makes any unknown model ID return 404 instead of 400 — silent breakage if your error handling branches on the old status code, so migrate Isaac calls before your next deploy.

  • What you can build this weekend

    Audio input covers WAV/MP3/FLAC up to ~21.8 minutes per clip, `enable_audio_in_video: true` reads a video's soundtrack, and outputs can be constrained with JSON Schema or regex; tool calls and sub-agents are supported. Context is ~32-37K multimodal tokens, with up to 4.7x faster chat and ~2x faster video. Try the demo at perceptron.inc/demo, grab a key at platform.perceptron.inc, `pip install perceptron`, or call it on OpenRouter.

  • Elsewhere: Qwen-Image-2.1 is unshippable

    Alibaba's new 7B open-weight image model — the #2 trending model on Hugging Face right now — does text-to-image, edits from up to 10 reference images, and outputs native transparent RGBA up to 2K, scoring 60.28 on Qwen's own bench vs Nano Banana 2.0's 59.82. The catch for builders: it dropped the previous version's Apache 2.0 license for a research-only license, so shipping it commercially now needs a separate agreement with Alibaba.

  • Elsewhere: MiMo-V2.6 tops the open board

    If you want a model you can actually ship, Xiaomi's MiMo-V2.6-Pro (1.02T params, 42B active, MIT license) is now first among open-weight models on the Artificial Analysis index at 46, tied with Grok 4.7, with a 1M-token context. The Flash variant runs $0.14/$0.28 per million tokens on hosts, and a MiMo-V2.6-Distill-Qwen-9B fits a capable agent onto a single GPU.