Kilo Code Weekly Product Roundup
Extension Version 4.140.3
Welcome back to the weekly product roundup! This week brings two powerful new models from Z.AI and MiniMax, plus a batch of Agent Manager stability fixes that make parallel development smoother than ever.
New Models
GLM-4.7
Z.AI’s newest model is now live in Kilo. GLM-4.7 isn’t just an incremental update; it fundamentally changes how open-weight models handle complexity.
Two improvements stand out:
Enhanced Programming Capabilities: GLM-4.7 implements a “think before acting” mechanism that delivers more stable performance on complex agent tasks. Multi-language coding and terminal agent applications see substantial gains.
Stronger Reasoning: Mathematical and reasoning skills jumped significantly, hitting 42.8% on the HLE (”Human Last Exam”) benchmark—a 41% improvement over GLM-4.6 and surpassing GPT-5.1.
To try it, open Kilo, switch your model to Z.AI: GLM-4.7, and start building.
MiniMax M2.1
MiniMax continues their momentum with M2.1, now available in Kilo. Their previous model M2 was trending at #3 on our AI coding models leaderboard for Code mode, so we’re watching closely to see how M2.1 stacks up.
The benchmarks are impressive: M2.1 scores 74.0 on SWE-bench Verified, edging out DeepSeek V3.2 (73.1) and Kimi K2 (71.3). On VIBE-Web, which tests the ability to build functional web apps, M2.1 hit 91.5 — outperforming Claude Opus 4.5 and Claude Sonnet 4.5 in some categories.
Our team ran M2.1 through real-world scenarios:
Building an Express API from scratch
Fixing bugs in a Next.js/React game
Extending a Go file-serving app
Documenting a complex Rust CLI
The verdict? M2.1 crushed all four, in one shot.
Not sure where to start? Try Code Reviews with M2.1 as your reviewing model, to take advantage of its speed and efficiency.
Extension
Agent Manager Improvements
This release includes several fixes that make Agent Manager more reliable and efficient:
Reduced GPU Usage: Message virtualization significantly cuts GPU consumption when running multiple agents in parallel.
Better CLI Detection: Fixed CLI detection and Windows spawn issues. Windows users should see much smoother parallel agent experiences in the IDE.
Multi-Version Session Handling: Agent Manager now properly waits for pending CLI processes, ensuring terminals are available per worktree. No more race conditions when spinning up multiple agents.
Improved Telemetry: Better visibility into what’s happening across your parallel agents.
New Features
App Builder — Build live apps and sites from the web with a live preview. Like Lovable, but integrated into the Kilo platform. Deploy or export the code with one click when you’re ready.
Code Reviews — AI-powered review agent that automatically analyzes your pull requests the moment they’re opened or updated. Surfaces issues across performance, security, style, and test coverage.
Parallel Agents — Run multiple Kilo agents without conflict in both the IDE and CLI. Agent Manager in the extension lets you launch 1-4 agents on git worktrees, each working on different tasks simultaneously.
Community Contributions
Thank you to our community contributors this week!
#4693 (@keeganwitt) - Add Requesty Codestral to autocomplete provider models
#4692 (@mcowger) - Fix loss of synthetic auto model refresh, which was lost in the Roo merge.
The parallel agent improvements are huge. Running multiple agents on seperate worktrees without stepping on each other was genuinly one of the biggest pain points. The GPU usage reductions through message virtualization is clever, especially when you're spinning up 3-4 agents simulatneously. Curious how Agent Manager handles task prioritization when agents hit overlaping dependencies