black box retweeted
Latest News - $BBOX
@askBlackBox has begun building Zero, its own model based on Qwen3.8-27B.
The base uses open weights under Apache 2.0, supports a 262K context window, accepts image input, and allows thinking to be switched off per request. Training starts with supervised fine-tuning followed by a preference-tuning run, using open datasets and examples generated by open-weight models. No user chats or outputs from closed models are included, consistent with Blackbox’s policy of not retaining prompts. Zero is being tuned for direct answers, skill following, tool use in the Code agent, and research. It will follow the same rules as other models on the platform: no prompts kept, history encrypted on the device, and existing hard lines maintained.
Once ready, Zero will be available in the app, API, and MCP, with evaluation results posted as runs complete. Blackbox provides private and unrestricted access to frontier models on Solana. The project marks the start of an in-house model built entirely from open sources.
For all the latest Solana tech news, check out techterminal.info/news. We track and cover updates from over 200+ Solana tech projects in one place.
Introducing Zero, Blackbox's own model. We've begun building it.
Base: Qwen3.8-27B. Open weights under Apache 2.0, 262K context, image input, and thinking you can switch off per request.
Training: supervised fine-tuning, then a preference-tuning run. The data is open datasets plus examples generated by open-weight models. Not one user chat, because Blackbox never kept any, and nothing from closed models.
What we're tuning it for: straight answers, following your skills, tool use in the Code agent, and research.
Same rules as every model on Blackbox: no prompts kept, history encrypted on your device, and the hard lines stay.
When it's ready, Zero goes into the app, the API and MCP. We'll post evals as runs finish.




