A million-token context is only useful if a model can act on what it reads. The notable shift here is the combination of long-context processing with tool-oriented work: coding, research, and visual tasks are treated as connected, multi-step workflows rather than isolated prompts. The trade-off is that “open-weight” does not mean lightweight: this is a 2.8T-parameter model.
Key Capabilities
- Long-context agentic work: The 1M-token window is intended for tasks that span large codebases or many documents, while tool use lets the model search, calculate, and iterate instead of relying only on a single response.
- Sparse MoE computation: Of 2.8T total parameters, 104B are active per token. This reduces per-token computation relative to activating the full model, but does not remove the substantial memory and infrastructure demands of hosting the weights.
- Text and image input: A vision encoder supports image understanding alongside text. The model card reports results across coding, visual reasoning, and document-based evaluations, including a 65.7 score on Terminal-Bench 4.0 and 92.7 on BenchCAD with tools. These are vendor-reported results, and scores across different benchmarks or evaluation setups are not directly comparable.
- Post-training for broad tasks: Moonshot describes large-scale joint reinforcement learning across hundreds of downstream domains. The intended payoff is more reliable execution across varied agent tasks, not just improved performance on one benchmark.
Who It Suits—and the Trade-off
A good fit for teams evaluating an open-weight model for long-context coding, multimodal analysis, or tool-driven workflows, especially when they can use hosted access or dedicated inference infrastructure. Look elsewhere if you need a small model for ordinary local hardware, a text-only model, or independently verified benchmark comparisons: the full model is exceptionally large, and many reported evaluations are vendor-run or use different harnesses.
