The strongest open-weight model we measured: #1 open on SWE-bench Pro, ahead of GPT-5.5. Built on GLM-5.2, hosted in the US.
Heavy moves to Z.ai GLM-5.2 (744B MoE, ~40B active, 1M context). Highest open-weight score on SWE-bench Pro (62.1%), ahead of GPT-5.5; first open model past 80% on Terminal-Bench 2.1. Behind the Claude frontier, shown honestly on the same axis.
Heavy tracks the strongest open-weight model on SWE-bench Pro. Z.ai moved 58.4 to 62.1 in a single GLM-5.1 to GLM-5.2 step; re-baselining is how that reaches you without a migration.
Standard public coding benchmarks, latest published runs.
Dropstone Heavy is the tier for the hardest, longest-horizon work. It runs Z.ai's GLM-5.2, a 744B-param Mixture-of-Experts model (roughly 40B active per token) with a 1M-token context. It scores 62.1% on SWE-bench Pro, the highest of any open-weight model and ahead of GPT-5.5 at 58.6%, and is the first open model past 80% on Terminal-Bench 2.1 (81.0%). It does not catch the closed Claude frontier (Opus 4.8 at 69.2%, Fable 5 at 80.3%), and we show that gap on the same axis as the win. For the large majority of production coding, the best open model, US-hosted at a fraction of frontier-closed cost, is now a reasonable default.
Changes that span many files and hundreds of steps. GLM-5.2's 1M-token context and Terminal-Bench lead keep it on track where weaker open models drift.
Hand Heavy a failing test and a stack trace and let it read beyond the immediate file, edit, and re-run until green. Its margin shows on the harder instances.
Iterative command-line work, the workload Terminal-Bench was built to measure, where GLM-5.2 leads the open field.
Every Dropstone request is treated as if the model could be adversarial. The CLI requires explicit user approval before any action that writes to disk, runs a shell command, or fetches a URL. No model output is ever auto-executed.
Heavy runs on SOC 2-certified, US-based inference providers. Enterprise and commercial data is never stored and never trained on, regardless of any setting. Consumer sessions help improve the open model we run and can be turned off anytime in settings, as text only, never your raw images.
Heavy is built on GLM-5.2, an open-weight foundation model. Goldwasser et al. (2022) proved no party can prove a closed foundation model is free of embedded behaviors, including Anthropic for Claude and OpenAI for GPT. We say this out loud. The runtime is why model origin does not matter for your code.
Same approval gate, same US-hosted inference, same zero-retention guarantee across all three. Pick the smallest model that meets the task.
Low-latency agentic coding for edits, refactors, and high-throughput inline completion. Built on DeepSeek V4 Flash, hosted in the US.
The everyday workhorse, built for tool use and day-to-day coding. Built on Kimi K2.7 Code, hosted in the US.
The strongest open-weight model we measured: #1 open on SWE-bench Pro, ahead of GPT-5.5. Built on GLM-5.2, hosted in the US.
Fast is included on every plan. Pro and Heavy are unlocked on the Pro plan ($15/mo) and Max plan ($75/mo).
Install the CLI, authenticate, and start running approval-gated agentic workflows. No credit card to start.
