Alledged Gemini 3.0 Pre-Release Models Lithiumflow and Orionmist on LMArena (winbuzzer.com)

Two mysterious models—codenamed “lithiumflow” and “orionmist”—appeared on LMArena over the weekend and are widely believed to be pre-release checkpoints for Google’s Gemini 3.0 (mapping roughly to Pro and Flash variants). Community sleuthing links the names and behavior patterns to Gemini 3.0: lithiumflow reportedly lacks web grounding, while orionmist appears to have search-enabled grounding. Early testers praise a marked leap in multimodal capability—clean, complex graphics-as-code (high-quality SVGs), robust visual reasoning (e.g., reliably reading analog clocks), and strong single-shot conceptual coding—though some note inconsistent coding performance, possibly due to public throttling or “nerfing” of these test builds. If genuine, these sightings matter because they suggest Gemini 3.0 could push state-of-the-art in vision+reasoning and code-generation, and introduce integrated grounding for web retrieval—features that would directly challenge recent releases like OpenAI’s GPT-5-Codex and Anthropic’s Claude Sonnet 4.5. The timing dovetails with an AI Studio overhaul and a rumored October 22 announcement, signaling Google may be close to a public launch. For practitioners, improved SVG/code generation and stronger visual comprehension mean new possibilities for programmatic graphics, vision-based automation, and retrieval-augmented workflows; for the industry, it tightens the arms race over multimodal, grounded, enterprise-ready models.