Skip to main content

Gemini 4 Argon (Understanding)

Gemini 4 and generative media​

Google’s public Gemini 4 catalog on ai.google.dev lists one stable understanding model: gemini-4-argon. It does not generate images or video (output is text only).

For the usual “Gemini stack” users mean alongside Gemini 4:

TaskOpenTryOn registry idGoogle endpoint
Multimodal understanding (reasoning)gemini-4-argongemini-4-argon
Image generation / editnano-banana-2.1 (also nano-banana-2, nano-banana-2-lite)gemini-nano-banana-2.1, etc.
Video generation / editgemini-omnigemini-omni-1.1-flash

Video-to-image (poster / thumbnail from a clip) uses generate --video on Nano Banana 2, 2.1, or 2 Lite via the Interactions API. See Nano Banana.

Fairwind-approved defensive cyber workloads may use gemini-4-argon-cyber-permissive on understand (allowlist; default thinking HIGH).

Gemini 4 Argon (gemini-4-argon) is Google’s flagship reasoning multimodal model on the Gemini API (stable since Oct 2026). OpenTryOn exposes it on understand via the same GeminiUnderstandAdapter as Gemini 3.8 Flash, with Argon-specific flags.

Registry idgemini-4-argon
AuthGEMINI_API_KEY
SDKgoogle-genai>=2.30.1
TierPaid (not free tier)

When to use which Gemini understand model

TaskModel
Fast, cost-efficient multimodal QAgemini-3.8-flash
Hard reasoning, long outputs, agentic video on long clipsgemini-4-argon

CLI (multimodal)​

export GEMINI_API_KEY="..."

opentryon understand --model gemini-4-argon \
  --video runway.mp4 --prompt "List each look and call out fabric changes." \
  --video-processing agentic --thinking-level medium --dry-run

Long decode (>64K tokens)​

Multimodal (recommended): set a high --max-tokens (e.g. 131072). The adapter enables automatic_continuation on generateContent (SDK ≥2.30.1).

opentryon understand --model gemini-4-argon \
  --pdf spec.pdf --prompt "Extract every measurement table as JSON." \
  --max-tokens 131072 --thinking-level high

Text-only background job (Interactions API): --background polls a server-side interaction (up to 24h). Requires --max-tokens greater than 65536 and no --image/--video/etc.

opentryon understand --model gemini-4-argon \
  --prompt "Write a detailed technical spec for a sizing API." \
  --max-tokens 262144 --background --thinking-level high

Python​

from tryon.api.gemini import GeminiUnderstandAdapter

adapter = GeminiUnderstandAdapter(model="gemini-4-argon")
out = adapter.understand(
    video="lookbook.mp4",
    prompt="Summarise the collection.",
    video_processing="agentic",
    thinking_level="medium",
    max_tokens=131072,
)
print(out["text"])

MCP​

Tool: understand_gemini_4_argon. Restart MCP after upgrading OpenTryOn.

See also​