LoginSign Up
← Back to ai-assistants
Weekly · ai-assistants · Week of October 19, 2026

LobeHub runs agents on users' devices while Copilot's autofix reaches general availability

device-side-executionmodel-onboardingagentic-autofixmulti-userrelease-consolidation
O
Generated 1h agoDrawn from 6 products

The week in ai-assistants

Assistants are moving onto the user's own devices and into the codebase. LobeChat, now shipping as LobeHub, released v2.2.20 stable, letting server-side agent runs call models that exist only in a user's browser tab or desktop app. GitHub Copilot took agentic autofix for code scanning alerts to general availability. The question is where the model runs and what it is allowed to change.

Model supply is the second thread. Copilot added Gemini 4 Argon on arrival, Baseten served its third new model API in eight days, and Ollama kept adding model support to its MLX runner.

Leaders

GitHub Copilot shipped agentic autofix for code scanning alerts at GA, which explores relevant files and proposes fixes across the codebase instead of patching one line, and Gemini 4 Argon, generally available on arrival for complex coding, agentic and reasoning work. A week earlier Haiku 5.5 took the fast-task slot.

LobeChat rolled 419 merged PRs into v2.2.20: device model relay including local Ollama downloads, goal batch rollouts, and a task board, with Codex able to run on connected devices.

InvokeAI reached 7 alpha 3, which turns boards into shared spaces with owners and four access roles and adds Qwen-Image-2.1. The v7 rewrite began with autosaving projects and a single widget workbench, and added MySQL and MariaDB backends.

Baseten added Reflection's Beam-1.0, with a 1M-token context window, tool calling and tunable reasoning. It also launched web search as a server-side hosted tool through Grounded Inference and egress restrictions that limit deployments to an allow list of IP ranges and domains.

Wildcards

Ollama moved to v0.40.6 release candidates that fix model pulls behind proxies when the registry redirects to a CDN, after adding Nemotron 3.5 Super to the MLX runner.

opencode retired its separate beta repository as the V2 tree became the main dev branch, and its VS Code extension reached 1.0 by launching the V2 CLI.

Themes that compounded

  • Multi-user and permission models arriving in local-first tools: role-gated boards at InvokeAI.
  • Device-side execution at LobeChat and Ollama.
  • Rapid model onboarding at GitHub Copilot and Baseten.
  • Release plumbing as a product decision: opencode consolidating onto V2.

Watch this week

Watch whether Ollama's v0.40.6 candidates reach a stable tag, and whether InvokeAI's v7 alphas move to a beta. Baseten's cadence of a new model API every few days is the clearest repeatable pattern in the sector.