Open-Weight Models Enter Mainstream IDEs: GLM-5.2/ZCode and Kimi K2.7 in Copilot
Key Questions
What capabilities does ZCode offer as the IDE for GLM-5.2?
ZCode provides persistent context, mobile remote access, and Git integration as the official agentic IDE for the GLM-5.2 model.
How does GLM-5.2 perform against GPT-5.5 on benchmarks and cost?
GLM-5.2 beats GPT-5.5 at 1/6 the cost on SWE-bench Pro (62.1 vs 58.6) and ties for strong results on Terminal-Bench Hard.
What does GitHub's addition of Kimi K2.7 Code signify?
It marks the first open-weight model in Copilot's selector, validating China's AI progress and pressuring providers toward similar integrations.
What updates are expected from Meta's Muse Spark release?
Muse Spark promises major gains in coding and agentic capabilities, with potential open-weight options noted alongside internal mandates for MetaCode.
Why is Kimi K2.7 Code considered a low-cost alternative?
As a 1T MoE model from Moonshot AI, it offers significantly reduced costs and is default-off for Enterprise/Business users in Copilot.
ZCode, the official agentic IDE for GLM-5.2, launched with persistent context, mobile remote, and Git integration. GLM-5.2 beats GPT-5.5 at 1/6 cost, MIT license, #2 on Code Arena frontend, ties Qwen3.7 Max at 50.8% on Terminal-Bench Hard, and leads on SWE-bench Pro (62.1 vs GPT-5.5 58.6). Business Insider article adds pricing comparisons and social media reactions. GitHub adds Kimi K2.7 Code as the first open-weight model in Copilot's model selector, default-off for Enterprise/Business. This signals a major shift: open-weight models becoming first-class citizens in major IDEs. Kimi K2.7 Code is a 1T MoE model from Moonshot AI, offering a low-cost alternative. The move validates China's AI capabilities and pressures other model providers to seek similar integrations. New: Meta's Muse Spark (Watermelon) update promises big coding/agentic gains, potentially catching GPT-5.5; open-weight possibility noted. New: Meta internal mandate to use MetaCode over third-party tools, targeting 65-80% AI-generated code, signals enterprise adoption of in-house open-weight models.