Pixtral 12B on real-time video streams
Pixtral 12B · Mistral AI · VLM · Open-weight
Pixtral 12B is Mistral AI’s first multimodal model, pairing a 12B parameter language backbone with a purpose-built 400M parameter vision encoder trained to handle images at their native resolution. It reads interleaved sequences of images and text without resizing everything to a fixed square, which keeps small details and unusual aspect ratios legible. Pixtral 12B is not currently in Overshoot's live model catalog. When a model like this is available through the API, it answers questions about a live WebRTC stream through an OpenAI-compatible chat-completions request.
Pixtral 12B is not currently in the live Overshoot model catalog (verified against the live model catalog on 2026-07-14). Availability changes over time - query GET /v1beta/models for the current list.
- Developer
- Mistral AI
- Parameters
- 12B + 400M vision encoder
- Context window
- 128K tokens
- License
- Apache 2.0
- Released
- Sep 2024
- Inputs
- Text, images, video frames
- Overshoot availability
- Not in live catalogas of 2026-07-14
What Pixtral 12B is good at
Pixtral 12B’s native-resolution vision encoder makes it a dependable choice for tasks where small visual details matter: reading text off a whiteboard or receipt, parsing a chart or table, or answering questions about a specific region of a busy frame. It also handles interleaved multi-image inputs cleanly, so a prompt can mix several frames with surrounding text.
As a fully open, Apache 2.0 model, Pixtral 12B is a solid open baseline to compare against when evaluating whether a heavier or closed model is actually needed for a given task. It is not a frontier reasoning model, but it is consistent and cheap to run at scale.
- OCR and document reading at native image resolution
- Chart and table interpretation
- Multi-image prompts mixing several video frames
Pixtral 12B and Overshoot's live-video workflow
Pixtral 12B is not currently in Overshoot's live model catalog, which GET /v1beta/models confirms at any time. Live footage on Overshoot works the same way for every model that is in the catalog: publish a WebRTC video track to a Stream, and Overshoot retains 600 seconds of frame history behind it. A chat-completions request then references that footage with an ovs:// URL, anchored to the latest frame, a timestamp, or a recent segment sampled at 1 fps by default.
If a model with Pixtral 12B's profile were serving those requests, answers would stream back over SSE with low time to first token, and a shared thread_id across requests would keep the prompt cache warm, which matters for applications that repeatedly re-ask about the same stream.
Pixtral 12B versus Pixtral Large
Pixtral 12B and Pixtral Large share the same native-resolution vision encoder design, but Large pairs it with Mistral Large 2’s 123B parameter backbone for substantially deeper reasoning. Pixtral 12B is the pick for high-throughput, cost-sensitive workloads like OCR and chart reading, while Pixtral Large is the pick when a task needs frontier-level visual reasoning rather than accurate extraction.
Frequently asked questions
Can Pixtral 12B analyze live video?
Pixtral 12B handles interleaved images and text, so it can analyze video by reading sampled frames when self-hosted or served by another provider. It is not currently in Overshoot's live model catalog, so it cannot be referenced against an Overshoot WebRTC stream today. Check GET /v1beta/models for the current catalog.
Is Pixtral 12B open source?
Pixtral 12B is released under the Apache 2.0 license with fully downloadable weights and no usage restrictions, so it is free to self-host, fine-tune, and deploy commercially.
Is Pixtral 12B available on Overshoot?
Not currently. Pixtral 12B is not in Overshoot's live model catalog. The catalog changes over time; GET /v1beta/models returns the current list of available models.
What is Pixtral 12B best used for?
Pixtral 12B is well suited to OCR, chart and table reading, and general visual question answering where cost and throughput matter more than frontier reasoning depth. For harder reasoning tasks, Pixtral Large is the stronger option.