DeepSeek V4.1 Pro has appeared in FlashRSI’s public quick-start example. We found it named as the target model in a SearchQA command, paired with V4.1 Flash as the worker. 1 2
The command combines the SkillOpt improvement method with a target agent running in MiniAgent. Pro answers the benchmark questions; Flash is the separate worker model used to propose edits to the agent’s resources. 1
We recently covered DeepSeek’s Flash 1017 update, and an updated Pro is the next release we are watching for. On X, we wondered whether the wait could involve distillation from GPT-6 Astra. That is still our guess; this repository does not establish how V4.1 Pro was trained. 2
The README also contains benchmark tables, but its shared experiment settings name deepseek-v4-pro, rather than deepseek-v4.1-pro. Those are different identifiers, so we are not treating the tables as V4.1 Pro results. The finding here is the newer Pro name in a public example, without a launch date or demonstrated public access. 1
Sources & references
- FlashRSI: quick-start command and experiment settings ↗
Pinned revision a265336ac30ba61f24211abb884179bb6252fd2c, checked 1 November 2026. The quick-start names deepseek-v4.1-pro as Target and deepseek-v4.1-flash as Worker. The separate overall-results section names deepseek-v4-pro as its evolution model.
- Lumina: DeepSeek V4.1 Pro benchmark-setup finding ↗
1 November 2026. Our original finding and supplied capture. The GPT-6 Astra distillation idea is Lumina’s speculation, not a claim established by the repository.
