Vision models on Overshoot

Amazon Nova Pro · Multimodal · Closed API

Amazon Nova Pro is Amazon's balanced frontier multimodal model with a 300K-token context window. It is not currently in Overshoot's live model catalog; check GET /v1beta/models for current availability.

Claude Haiku 4.5 · Multimodal · Closed API

Run Claude Haiku 4.5 on live video with Overshoot: publish over WebRTC, reference frames with ovs:// URLs, and get fast answers through a compatible API.

Claude Opus · Multimodal · Closed API

Query Claude Opus on live video through Overshoot: publish a WebRTC stream, reference frames with ovs:// URLs, and get deep visual reasoning back over SSE.

Claude Sonnet 4.5 · Multimodal · Closed API

Use Claude Sonnet 4.5 on live video with Overshoot: publish over WebRTC, reference frames with ovs:// URLs, and call the same OpenAI-compatible endpoint.

Gemini 3 Flash · Multimodal · Closed API

Use Gemini 3 Flash on live video with Overshoot: publish over WebRTC, reference frames with ovs:// URLs, and call the same OpenAI-compatible endpoint.

Gemini 3 Pro · Multimodal · Closed API

Query Gemini 3 Pro on live video through Overshoot: publish a WebRTC stream, reference frames with ovs:// URLs, and get long-context reasoning back over SSE.

GPT-4o · Multimodal · Closed API

GPT-4o is OpenAI's natively multimodal omni model. Learn its vision capabilities and its current availability in Overshoot's live model catalog, which changes over time.

GPT-5 · Multimodal · Closed API

Query GPT-5 on live video through Overshoot: publish a WebRTC stream, reference frames with a chat-completions request, and stream the answer back.

GPT-5 mini · Multimodal · Closed API

Use GPT-5 mini on live video with Overshoot: publish over WebRTC, reference frames with an ovs:// URL, and get answers through an OpenAI-compatible endpoint.

GPT-5.4 nano · Multimodal · Closed API

Run GPT-5.4 nano on live video streams with Overshoot: WebRTC publishing, ovs:// frame references, and OpenAI-compatible chat completions at low cost.

Grok 4 · Multimodal · Closed API

Grok 4 is xAI's frontier reasoning model with native vision. Learn its strengths on frames and diagrams and its current availability in Overshoot's model catalog.

SenseNova-V · VLM · API

SenseNova-V is SenseTime's hosted vision flagship, strong at multi-frame video comprehension. It is not currently in Overshoot's live model catalog; see how it fits real-time video.

Aya Vision 32B · VLM · Open-weight

Aya Vision 32B covers 23 languages for captioning and visual QA. It is not currently in Overshoot's live model catalog; check GET /v1beta/models for current availability.

Command A Vision · VLM · Open + API

Command A Vision reads charts, tables, and documents for enterprise use. It is not currently in Overshoot's live model catalog; check GET /v1beta/models for current availability.

DeepSeek-VL2 · VLM MoE · Open-weight

DeepSeek-VL2 brings sparse mixture-of-experts vision-language modeling with strong OCR and grounding. It is not currently in Overshoot's live model catalog; check GET /v1beta/models.

ERNIE 4.5 VL · VLM MoE · Open + API

ERNIE 4.5 VL is Baidu's heterogeneous MoE vision model for charts, documents, and scenes. See how it applies to real-time video and check Overshoot's live model catalog.

Gemma 3 27B · Multimodal · Open-weight

Gemma 3 27B pairs a SigLIP vision encoder with 140+ language support. See how it applies to real-time video and where it sits next to the Gemma 4 generation.

Gemma 4 26B-A4B · Multimodal MoE · Open-weight Hosted on Overshoot

Gemma 4 26B-A4B answers questions about live WebRTC video on Overshoot in about 200ms, with a sparse 4B-active design built for high-volume streaming.

Gemma 4 31B · Multimodal · Open-weight Hosted on Overshoot

Run Gemma 4 31B on live video with Overshoot: stream over WebRTC, query frames with an OpenAI-compatible API, and get answers in about 200ms.

GLM-4.5V · VLM MoE · Open-weight

GLM-4.5V is Zhipu AI's open MoE vision-language model with a switchable thinking mode. See how it applies to live video and check Overshoot's current model catalog.

HunyuanVL · VLM · Open + API

HunyuanVL is Tencent's open-weight vision flagship with strong OCR and bilingual grounding. See its capabilities and its current status in Overshoot's model catalog.

InternVL3 78B · VLM · Open-weight

InternVL3 78B pairs native multimodal pretraining with a Qwen2.5 backbone for flagship-class visual reasoning. See its current status in Overshoot's model catalog.

Llama 3.2 Vision · VLM · Open-weight

Llama 3.2 Vision is Meta's adapter-based vision model in 11B and 90B sizes. See how it handles video frames and its current Overshoot availability.

Llama 4 Maverick · Multimodal · Open-weight

Llama 4 Maverick is Meta's flagship multimodal MoE model with 17B active parameters. See how it handles video frames and its current Overshoot availability.

Llama 4 Scout · Multimodal · Open-weight

Llama 4 Scout is Meta's multimodal MoE model with a 10M-token context window. See how it handles live video frames and its current Overshoot availability.

LLaVA-OneVision · VLM · Open-weight

LLaVA-OneVision trains one recipe across images, multi-image sets, and video. See what it is good at and its current availability on Overshoot's API.

MiniMax-VL-01 · VLM · Open + API

MiniMax-VL-01 uses lightning attention and a 4M-token window built for long-video understanding. See its strengths and current Overshoot availability.

Molmo 72B · VLM · Open-weight

Molmo 72B is the Allen Institute for AI's pointing-focused open vision model. See how it fits live video work and its current Overshoot availability.

NVLM-D 72B · VLM · Open-weight

NVLM-D 72B is NVIDIA's open research vision-language model with strong OCR and dynamic tiling. See its capabilities and its current status in Overshoot's model catalog.

Pixtral 12B · VLM · Open-weight

Pixtral 12B is Mistral AI's open multimodal model with a native-resolution vision encoder. See how it fits live video and its status in Overshoot's model catalog.

Pixtral Large · VLM · Open-weight

Pixtral Large is Mistral AI's frontier open-weight multimodal model. See its visual reasoning strengths and its current status in Overshoot's model catalog.

Qwen2.5-VL 72B · VLM · Open-weight

Qwen2.5-VL 72B is Alibaba's prior-generation flagship VLM with precise grounding and document parsing. See its strengths and its status in Overshoot's catalog.

Qwen3-VL 235B-A22B · VLM MoE · Open-weight

Qwen3-VL 235B-A22B is Alibaba's MoE flagship VLM for hard visual reasoning and long video. It is not currently in Overshoot's live model catalog; see how it fits real-time video.

Qwen3-VL 32B · VLM · Open-weight

Qwen3-VL 32B is Alibaba's dense flagship VLM for video and long-document understanding. It is not currently in Overshoot's live model catalog; see how it fits real-time video.

Qwen3.6 27B · Multimodal · Open-weight Hosted on Overshoot

Run Qwen3.6 27B on live video with Overshoot: publish over WebRTC, reference frames with ovs:// URLs, and stream answers back in about 200ms.

Qwen3.6 35B A3B · Multimodal MoE · Open-weight Hosted on Overshoot

Run Qwen3.6 35B A3B on live video with Overshoot: publish over WebRTC, query frames with an OpenAI-compatible API, and stream answers in about 200ms.

Step-3 · VLM MoE · Open-weight

Step-3 is StepFun's hardware-aware MoE vision model built for cheap decoding at scale. It is not currently in Overshoot's live model catalog; see how it fits real-time video.

FastVLM 7B · On-device · Research

FastVLM 7B is Apple's on-device vision model, built for near-instant time to first token on high-resolution frames. See how it fits real-time video workloads.

Granite Vision 3.2 · Doc VLM · Open-weight

Granite Vision 3.2 is IBM's compact document vision model for charts, tables, and forms. See its capabilities and its current status in Overshoot's live model catalog.

Kimi-VL A3B · VLM MoE · Open-weight

Kimi-VL A3B activates just 2.8B parameters for efficient long-video reasoning. See Moonshot AI's model capabilities and its status in Overshoot's model catalog.

LFM2-VL 3B · Edge VLM · Open-weight

LFM2-VL 3B is Liquid AI's edge-tuned vision model with native-resolution encoding. See how it fits real-time video and its current Overshoot availability.

MiniCPM-V 4.5 · Small VLM · Open-weight

MiniCPM-V 4.5 is OpenBMB's compact video-and-OCR vision-language model. See how its 3D-Resampler fits live video and its current Overshoot availability.

Mistral Small 3.1 · Multimodal · Open-weight

Mistral Small 3.1 is Mistral AI's 24B open-weight model that adds vision to the Small line. See how it fits live video and its current Overshoot availability.

Moondream 2 · Tiny VLM · Open-weight

Moondream 2 is M87 Labs' tiny open vision model with detection and pointing. See its strengths for real-time video and its current status in Overshoot's model catalog.

Phi-3.5 Vision · Small VLM · Open-weight

Phi-3.5 Vision is a 4.2B multi-frame image model from Microsoft. See its strengths for live video work and its current status in Overshoot's model catalog.

Phi-4 Multimodal · Small VLM · Open-weight

Phi-4 Multimodal is Microsoft's compact 5.6B vision, speech, and text model. See its strengths for live video and its current status in Overshoot's model catalog.

Qwen2.5-VL 7B · Small VLM · Open-weight

Qwen2.5-VL 7B is Alibaba's efficient open-weight VLM for OCR and grounding on video frames. It is not currently in Overshoot's live model catalog; see how it fits real-time streaming workflows.

SmolVLM2 2.2B · Tiny VLM · Open-weight

SmolVLM2 2.2B is Hugging Face's laptop-scale, video-capable vision-language model. It is not currently in Overshoot's live model catalog; see how tiny VLMs fit real-time video.

CogVLM2 · VLM · Open-weight

CogVLM2 pairs a Llama 3 8B backbone with deep visual-expert grounding for GUI and document reading. It is not currently in Overshoot's live model catalog; check GET /v1beta/models.

Emu3 8B · Any-to-any · Open-weight

Emu3 8B is BAAI's any-to-any transformer for text, image, and video understanding and generation. It is not currently in Overshoot's live model catalog; check GET /v1beta/models.

Florence-2 · Vision foundation · Open-weight

Florence-2 is Microsoft's prompt-based vision foundation model for captioning, detection, and OCR, a natural fit for per-frame tasks in real-time video pipelines.

Holo 3.1 35B A3B · Computer-use · Open-weight Hosted on Overshoot

Run Holo 3.1 35B A3B on live screens with Overshoot: publish over WebRTC, query frames via an OpenAI-compatible API, and get UI-grounded answers in about 200ms.

Holo3 35B A3B · Computer-use · Open-weight Hosted on Overshoot

Run Holo3 35B A3B, H Company's open computer-use MoE model, on live screen shares with Overshoot's WebRTC streams and OpenAI-compatible chat API.

Janus-Pro 7B · Unified · Open-weight

Janus-Pro 7B unifies visual understanding and image generation in one transformer. See DeepSeek's model capabilities and its status in Overshoot's model catalog.

PaliGemma 2 28B · Vision-LM · Open-weight

PaliGemma 2 28B is a prefix-LM vision model for captioning, detection, and segmentation transfer. See how it fits live video and its status in Overshoot's catalog.

QVQ-72B · Visual reasoning · Open-weight

QVQ-72B is Alibaba's experimental visual reasoning model that works through problems step by step. See its strengths and its current status in Overshoot's catalog.

Skywork-R1V 38B · Reasoning · Open-weight

Skywork-R1V 38B is a multimodal reasoning model with chain-of-thought transfer, strong on visual math and logic. It is not currently in Overshoot's live model catalog.

UI-TARS 72B · Computer-use · Open

UI-TARS 72B is ByteDance's native computer-use model, emitting GUI actions from screenshots. It is not currently in Overshoot's live model catalog; see how it fits screen streaming.