It's a been a great 6 months to be using local AI as the performance delta has, on average, been very low for classic LLMs, with R1 typically being at or near SOTA, and smaller models consistently getting better and better benchmarks.

However, the below are all things where there has been a surprising lag between closed systems' release dates and the availability of high quality local alternatives

1. A voice mode that is on par with Chat Gpt. Most all the pieces are in place to have something akin to 4o with voice. Sesame, Kyutai, or Chatterbox for TTS, any local model for the LLM, decent STT is, I think, also a thing already. We just need the parts put together in a fairly user-friendly, fast streaming package.

2. Local deep research on the level of o3's web search. o3 is quite amazing now in its ability to rapidly search several web pages to answer questions. There are some solutions for local llms but none that I've tried seem to be fulfilling the potential of web search agents with clever and easily customizable workflows. I would be fine with a much slower process if the answers were as good. Something like Qwen 235b I believe could do a great job of being the foundation of such an agent.

3. A local visual llm that can reliably read any human-legible document. Maverick is quite good but not nearly as good as Gemini Pro or Chat GPT at this.

What else am I forgetting about?