Flash Onyx 2.2 has landed, and this update is all about making local AI coding assistance feel less like talking to a chatbot and more like working with a terse senior engineer who actually read the codebase before opening their mouth.
What's New in 2.2
The headline feature: Flash Onyx 2.2 builds on gemma4 with an enhanced system prompt that turns the model into what its creator calls an "engineering agent." The workflow is refreshingly simpleβread the files, make the change, run the thing, report back in as few words as the truth allows. No fluff, no apologetic hedging.
Local-First Philosophy
Flash itself positions itself as a local-first agent shell. That means Onyx 2.2 runs entirely on your hardwareβthe post mentions it working on systems with just 16 GB of RAM. For developers who have privacy concerns about shipping code to cloud APIs, or who work in air-gapped environments, this is increasingly viable without breaking the bank.
The "Game Feel" Angle
The title hints at something interesting: "teaching a local model law and game feel." The system prompt engineering approach suggests they're not just asking gemma4 to generate codeβthey're shaping how it responds to match developer expectations around concision, accuracy, and actionability.
Key Takeaways
- Flash Onyx 2.2 uses gemma4 as its base model with custom system prompts for agentic behavior
- Runs locally on consumer hardware (confirmed working at 16 GB RAM)
- Designed for a tight feedback loop: read files, make changes, execute, report errors concisely
- Part of the broader local-first/LLM movement pushing back against cloud dependency
The Bottom Line
The gap between local model capability and practical developer tooling keeps shrinking. Flash Onyx 2.2 isn't trying to replace Claude or GPT-4βit's building something leaner for devs who'd rather keep their code on their own machines.