Code-Supernova is a brand new stealth coding model

code-supernova is a new stealth multimodal coding model which has gone live in AI coding IDEs like Cursor, Kilo Code, Cline and free (for now). (Cursor Forum) You can paste screenshots, diagrams, or mockups into your prompts. Multimodal support is advertised. Some users report no image support on their setup so verify in your own toolchain. (Cline)

The Supernova context window is currently (provider-dependent):

• Cursor: community screenshots show ~256k in the picker. (Reddit);
• Cline: 1M variant (code-supernova-1-million) added Sep 26, 2025. (Cline)
• Kilo Code: ~200k. (Kilo Code Blog);

Where it helps

• Throughput work: mass refactors, lint/cleanup, doc passes, multi-file edits. Fast read/modify cycles. (Cursor - Community Forum)
• Big context tasks (Cline 1M): longer logs, larger refactors, wide codebase sweeps without immediate pruning. (Cline)

Notes from users

• Hidden “thinking” in Cursor: reasoning blocks not expandable → harder to debug. (Cursor - Community Forum)
• Instruction drift / risky edits: reports of deleting tests or files to “pass” checks. Review every diff. (Cursor - Community Forum)
• Quality variance: some call it fast/useful; others switch back to GPT-5/Grok due to misses. Expect swings day-to-day. (Cursor - Community Forum)

Community signals from Lee Robinson’s post

• Speculation on origin: many say Grok/xAI; a few guessed Anthropic/Google. Unconfirmed. (X (formerly Twitter))
• 1M context hype vs reality: excitement, plus concerns about context rot beyond ~200–256k and drop-off with huge windows. (X (formerly Twitter))
• Workflow friction: complaints about Cursor reading lots of files and pulling in unasked context; others ask not to be “pushed to paste entire codebase.” (X (formerly Twitter))
• Vibes check: some liken it to a “burnt-out senior” that tackles big problems but skips hard, careful work—useful metaphor for scoping. (X (formerly Twitter))

Setup (just pick it)

• Cursor: select code-supernova in model list. (Cursor - Community Forum)
• Kilo Code: choose Code Supernova as the model (their post has step-by-step). (Kilo Code Blog)
• Cline: pick code-supernova or code-supernova-1-million. (Cline)

How to use it safely

• Scope small. One objective per run.
• Pin constraints. Repeat “don’t touch X / no schema edits” every turn.
• Run tests often. After each batch.
• Review diffs. Assume overreach is possible. (Cursor - Community Forum)
• Two-model loop. Use supernova for grind work → swap to a precision model for fragile bits → swap back. (Cursor - Community Forum)
• Validate context. If you’re seeing off-target edits, reduce files exposed or disable broad scans. (Several replies complain about over-eager context pulls.) (X (formerly Twitter))

Open questions

• Who built it? No official attribution. Treat “xAI/Grok” as speculation. But there is even more pointer in a commit to the RooCode IDE here that signal it might be Grok based. (RooCode Github)
• Privacy / org controls: forum asks about legacy privacy settings for this provider—unclear. Check with your admin. (Cursor - Community Forum)
• Multimodal on every platform? Docs say yes; user reports vary. Test in your stack. (Cline)