← sources

My local model is now fully integrated with the devices around me and has access to the same capabilities as Cowork or Code, at a fraction of the token price.

source: reddit (reddit:ollama)published 2026-10-29 13:30ingested 2026-10-30 02:40

retained text

some of you probably remember my earlier posts about aletha codex. since then i've built out a full coding environment with terminal access, git, testing, and project execution alongside cowork, while keeping persistent memory independent of whichever model you're running. Through smaller local embedding model handles retrieval across chats and model switches.

i also benchmarked Aletha against opencode, Ader, and cline using gpt-oss:120b across 24 coding runs each. Aletha finished 24/24, Aider 19/24, OpenCode 18/24, and Cline 11/24. Aletha averaged 78k prompt tokens per solved task versus OpenCode's 150k, although Aider was more token-efficient but less effective. it's a small benchmark, but getting those results out of the harness i've built was pretty great, love seeing the work payoff. the video shows the actual system working, including device control, Cowork, and Code. along with some small explanations to how it works ect. I also integrated Cowork and Code into the Vault, restricting them to local models within an AES256 encrypted workspace. I know a lot of people here enjoy experimenting with abliterated models, myself included, so I wanted to create an environment where you have more privacy and control over whatever you're working on. 

i genuinely think people underestimate how valuable these open models are, even if they're still behind the frontier labs. we don't know what restrictions could look like in the future, or whether access might eventually be limited. Thats why I think it’s important to be able to rig up your own shit and have it be reliable and capable of working offline. Let me know what you guys think and if you want more detailed explanations feel free to ask about any of it. The hardest part was the agent harness, developing the memory embedded model, and giving it the same kind of capabilities that can work offline anywhere. Alot to unpack in this post I know but just coming to hear about what you guys think about the rig.

document fetched 2026-10-30 00:44 · canonical https://reddit.com/r/ollama/comments/1xik0vc/my_local_model_is_now_fully_integrated_with_the