Models¶
The Macaron hosted API serves Macaron-V1.1, the original Macaron-V1 lineup, and GLM base models. Macaron-V1.1 combines a GLM-5.3 base with four LoRA specialists for Chat, Agent, Coding, and Generative UI. For release details and the original V1 architecture, see What is Macaron.
Hosted lineup¶
| Model / API ID | Aggregate params | Base | Specialists | Use |
|---|---|---|---|---|
macaron-v1.1-venti (Macaron-V1.1) |
752B | 744B GLM-5.3 | 4 × 2B LoRA (Chat, Agent, Coding, Generative UI) | Long-horizon coding, complex office tasks, chat, and GenUI |
macaron-v1-venti |
748B | 744B GLM-5.2 (frozen) | 4 × 1B LoRA (L0 Chat, L1 Agent, L2 Coding, L3 GenUI) | Flagship; chat, agent, coding, GenUI via routing |
macaron-v1-tall |
50B | Qwen3.6-35B-A3B (frozen) | 4 × 1B LoRA (same four-specialist design) | Local / lower-latency deployment |
macaron-v1-coding-venti |
748B | 744B GLM-5.2 | coding LoRA merged into base (single-specialist) | Coding-specialized; no routing overhead |
glm-5.3 |
744B | GLM-5.3 | none (base only) | Base model used for Macaron-V1.1 post-training |
glm-5.2 |
744B | GLM-5.2 | none (base only) | Base model without MoL specialists |
Use the model ID shown in the dashboard when making API requests. Names in code font above are API IDs. The hosted API also lists macaron-v1.1 and macaron-v1.1-coding-venti.
For Macaron-V1.1 thinking, sampling, and output-length parameters, see Parameters.
Routing¶
In the original V1 lineup, the two MoL-served variants (macaron-v1-venti, macaron-v1-tall) route each user turn through a three-stage loop: route (L0 classifies the request into one specialist under a 24-token budget), answer (the chosen specialist responds), summary (the specialist emits a ≤192-token summary stored server-side as shared context). macaron-v1-coding-venti skips routing because its coding LoRA is merged into the base. The GLM-5.2 and GLM-5.3 base models have no MoL specialist routing.
Macaron-V1.1¶
Macaron-V1.1 is a 752B-parameter agent model post-trained from GLM-5.3: a 744B base plus four 2B LoRA specialists. It upgrades the base from GLM-5.2 to GLM-5.3. Using Mint Recursive, Mind Lab completed the iteration from Macaron-V1 to V1.1 in two weeks.
The release focuses on completing long-horizon coding tasks in fewer steps and tokens, and on choosing when and how to use tools in complex office workflows. ChatBench and UI4A-Bench also expand to V2 with more real-world cases.
The release introduction includes a Chat experiment across 16 dimensions and a UI4A comparison across 60 tasks. In the UI4A comparison, first-attempt delivery success rose from 71.67% (43/60) for GLM-5.3 to 96.67% (58/60) for Macaron-V1.1. See the V1.1 introduction and case studies for the reported improvements, regressions, and access options.
Pricing¶
Model unit prices change as models evolve. The dashboard's Models & Pricing remains the source for current rates: mainland China (mintcn.macaron.xin, prices in RMB) or international (mint.macaron.im, prices in USD). For Training API, Inference API, and LLM as a judge service billing and solution pricing, see Mint Recursive → Service Pricing.
Self-hosted base models¶
The lineup above contains hosted Macaron and GLM models. If you are self-hosting Mint Recursive and serving Qwen3 community base models on your own endpoint, see the self-hosted models reference for the Qwen3 lineup. The two surfaces are distinct: the hosted API uses /v1/chat/completions; the self-hosted surface serves Qwen3 base models over /oai/api/v1.