Skip to content

v0.0.9 — Decision models, robot-act / detect / parse, Gemini 4 Argon and Omni 1.2

Latest

Choose a tag to compare

@kailashahirwar kailashahirwar released this 11 Nov 08:24
· 4 commits to main since this release
2f96092

OpenTryOn v0.0.9 Release Notes

Decision models, three new services, Gemini 4 Argon and Omni 1.2, Kimi K3.1 Pro

Release Date: 11 November 2026

OpenTryOn v0.0.9 adds 47 model ids and three service categories: robot-act, detect and parse. 210 models, 215 MCP tools (210 models + 5 discovery / key / planner), up from 163 models / 168 tools. Sixteen services: vton, generate, edit, understand, video-generate, audio-to-video, generate-3d, bg-remove, tts, decide, embed, classify, measure, robot-act, detect, parse.

What's New

Decision models (decide)

decide grows from 7 to 27 models. They all share one question schema (noul / choice / score) and return calibrated probabilities instead of text, so you can swap models without changing your questions.

CLI model Path What it is
pplx-decider Perplexity Decisions API (PERPLEXITY_API_KEY) Decider v1.2 (27B) / Flash v1.2 (4B); images; optional web grounding
pplx-decider-local your SGLang server Open perplexity-ai/pplx-decider-v1.1-27b
clef Cloudflare Workers AI (CLOUDFLARE_API_TOKEN + CLOUDFLARE_ACCOUNT_ID) Clef (27B) / Clef-flash (9B); images
clef-local your SGLang server Open Cloudflare/clef
kev-openrouter OpenRouter (OPENROUTER_API_KEY) Jared Palmer's Kev 4B (no first-party API exists)
kev-local the Kev repo's kev.serve Kev 0.8B to 27B; MLX on Apple Silicon
autotrust-jev-local your vLLM serve.sh AutoTrust JEV-35B / 27B-VL (no hosted API exists)
d3-turbo … d3-edge (+ -local) vLLM Semantic Router (DECISION_API_KEY) Decision 3.0, 0.6B to 27B; adds box and track questions
mercury-decide Inception (INCEPTION_API_KEY) Text + images

Understand

CLI model What it is
gemini-4-argon Google's flagship reasoning model: agentic video processing, long decode, background jobs (GEMINI_API_KEY)
gemini-4-argon-cyber-permissive Allowlisted variant for defensive cyber work
kimi-k3.1-pro Moonshot's K3 successor, 1M context, image + video; --kimi-model kimi-k3.1-pro-highspeed for the faster variant
moondream-3.1, moondream-3-preview, moondream-3.1-local Query and caption (MOONDREAM_API_KEY, or local Photon)
mercury-3 Inception's diffusion LLM, text-only, 1M context
beam, flint-local / flint-openrouter, sage-a1.5-local, ripple-65b-openrouter / ripple-65b-local Text-only agent and coding models, named-model-only

Image and video

CLI model Service What it does
flux3-max video-generate FLUX 3 Max (BFL); flux3-video gains --version latest|max
kling-v4 video-generate Kling 4.0
gemini-omni video-generate Now defaults to Gemini Omni 1.2 Flash: up to 15s with native audio, new --duration
seedream --model-version seedream-5-0-flash generate + edit Seedream 5.0 Flash, the faster, cheaper 5.0 tier
nano-banana-2 / -2.1 / -2-lite with --video generate Poster or thumbnail from a video clip
cosmos3-nano-t2i, cosmos3-nano-v2v generate, video-generate Cosmos 3 Nano certified NIM (self-hosted)

New robot-act, detect and parse services

CLI model Service What it does
dream-1, dream-1-flash robot-act Dream Labs World Action Model (DREAM_API_KEY)
cosmos3-nano-droid robot-act Cosmos 3 Nano-DROID policy (certified NIM)
moondream-3.1, moondream-3-preview, moondream-3.1-local detect Boxes, points and SVG segmentation masks
nemotron-parse, nemotron-parse-local parse NVIDIA document OCR (NIM hosted or local)

Embed

CLI model What it is
nemotron-3.5-embed-1b, nemotron-3.5-embed-1b-local NVIDIA text + document-image embeddings

Breaking changes

  • google-genai>=2.30.1 is now required (was >=1.52.0). Gemini Omni, Gemini 4 Argon background jobs and Nano Banana video-to-image use the Interactions API, which the 1.x SDK lacks. pip install -U opentryon upgrades it.
  • gemini-omni defaults to gemini-omni-1.2-flash. Pass --omni-model gemini-omni-1.1-flash to keep 1.1 until Google routes it to 1.2 on 3 December 2026.
  • No model ids were removed. New services and providers are opt-in and need their own keys or a self-hosted server.

Fixes

  • Seedream: seedream sent seedream-5-0-pro-260628, which BytePlus ModelArk does not list. It now sends the published dola-seedream-5-0-pro-260628, and 5.0 Lite and 4.5 use their dated ids. The --model-version names you pass are unchanged.

Install / Upgrade

pip install -U opentryon
pip install -U "opentryon[local]"

From source:

git checkout v0.0.9
pip install -e .

Quick examples

opentryon decide --model pplx-decider --questions questions.json --image dress.jpg
opentryon decide --model clef --clef-model clef-flash --questions questions.json --state "Red satin gown"
opentryon understand --model gemini-4-argon --video runway.mp4 --prompt "List every look and its fabric"
opentryon understand --model kimi-k3.1-pro --image look.jpg --prompt "Describe the outfit"
opentryon video-generate --model gemini-omni --prompt "Slow tracking shot down a runway" --duration 15
opentryon detect --model moondream-3.1 --image look.jpg --object "handbag" --skill segment
opentryon parse --model nemotron-parse --image size_chart.png

Docs

Notes

  • Restart MCP after upgrade so clients see the new tools.
  • TryOn Studio capability screens cover Image, VTON, Understand, Video, BG Remove, TTS and Detect. decide, embed, classify, measure, robot-act, parse, audio-to-video and generate-3d are available from the CLI and any MCP client, including Agent chat when you name the model.
  • Text-only models such as beam, flint-*, sage-a1.5-local and ripple-65b-* are outside the fashion/media scope and are named-model-only.
  • 210 models across 16 services → 215 MCP tools.

Links