
doing smart stuff @Token_Logic prev @summerfinance_ | @oasisdotapp | @aptiv
- Tweets1.9K
- Following1K
- Followers3.6K
- Likes2.2K

ALT LokalBot 0.10.1: what got faster since 0.9.2. A table of six measurements, before and now. Autocomplete at the end of a long draft: 257 ms before, 60 ms now (model time per suggestion, M4 Max). Idle main thread: about 5% busy before, 99.5% idle now (30 s sample at 10 ms). Notes for a 20-minute meeting: 4 model requests before, 1 or 2 now (hosted models such as Qwen 3.8 27B). Overnight review: 80 to 480 s before, 26 s now (one review of a day, measured live). Launch cleanup: 566 ms before, 196 ms now (180-day library pruned to 14 days). Transcription with Qwen3-ASR 0.6B: 27 times real time before, 116 times now (benchmark harness, M4 Max). lokalbot.com, free and open source, private work memory for your Mac. Sources: merged pull requests 129, 141, 152, 163 and 200 on github.com/stevyhacker/lokalbot.

ALT Infographic: "The little AI team inside your Mac." A robot next to a MacBook running LokalBot app using LLMs, with six AI model cards: Nemotron 3 (Core ML, 0.2 GB), Qwen3-ASR 1.7B (MLX 8-bit, 2.47 GB), LFM2.5 1.2B (GGUF Q4_K_M, 0.73 GB), Harrier 0.6B (GGUF Q8_0, 0.64 GB), Qwen3.5 4B (GGUF Q4_K_M, 2.74 GB) and Apple Vision (built into macOS). 6.8 GB, all local.

ALT Infographic: "The little AI team inside your Mac." A robot next to a MacBook running LokalBot app using LLMs, with six AI model cards: Nemotron 3 (Core ML, 0.2 GB), Qwen3-ASR 1.7B (MLX 8-bit, 2.47 GB), LFM2.5 1.2B (GGUF Q4_K_M, 0.73 GB), Harrier 0.6B (GGUF Q8_0, 0.64 GB), Qwen3.5 4B (GGUF Q4_K_M, 2.74 GB) and Apple Vision (built into macOS). 6.8 GB, all local.






























