Models

Every plan includes the launch model with no token meter. Models marked coming soon are not available at launch. Source data: models.json. Lineup updated 2026-11-10; it may change as better open-weight coding models ship.

DeepSeek V4.1 Flash-1017

• Id: deepseek-v4.1-flash
• Availability: at launch
• Maker: DeepSeek
• Size: 552B backbone MoE + 196B Engram memory (763B total per Hugging Face); 8B active in prefill, 16B in decode
• License: MIT
• Router role: Default model at launch; serves every request until a second model is added.
• Model card: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017

GLM-5.3-Flash (coming soon)

• Id: glm-5.3-flash
• Availability: coming soon, not at launch
• Maker: Z.ai
• Size: 320B total / 18B active; hybrid sparse + linear attention; multimodal
• License: MIT
• Router role: Added to the auto router for tool-heavy and vision requests when it ships.
• Model card: https://huggingface.co/zai-org/GLM-5.3-Flash

Router

Model id auto: At launch it sends every request to DeepSeek V4.1 Flash-1017. When GLM-5.3-Flash is added, auto will pick a model per request; you can always pin a model id instead.

Vendor-reported scores

We haven't published our own benchmarks yet. These are the model makers' self-reported scores; each number links to its source. Harnesses differ, so this is not a head-to-head comparison, and results through any provider can differ.

• DeepSeek V4.1 Flash-1017: model card. Self-reported by DeepSeek on the model card.
• GLM-5.3-Flash: model card. Self-reported by Z.ai (blog and Hugging Face eval results).

Agents: /docs/models.md is this page in markdown, or request it with Accept: text/markdown.