AI Agent - Aug 2, 2026

Macaron V1 API vs Local Deployment Access Guide

Quick answer

Use the hosted API to evaluate Macaron-V1-Venti without provisioning a 748B serving stack. Use open weights when you need infrastructure ownership, inspection, or local controls; Mind Lab specifically positions the 50B Macaron-V1-Tall for local deployment.

Do not treat these paths as interchangeable. Confirm model identity, region, endpoint, revision, license, hardware, pricing, rate limits, privacy, retention, reliability, and support for the path you choose.

Access paths

PathProvider-described availabilityMain verification burden
Overseas hosted APImint.macaron.im; release describes it as a Venti trial pathLive model ID, authentication, pricing, rate limit, region, privacy, retention, SLA
Mainland-China hosted APImintcn.macaron.xin; separate regional entry pointAccount eligibility, entity, terms, data location, network, pricing, support
Venti open weightsHugging Face Macaron-V1 collection748B infrastructure, license, base obligations, serving, quantization, security
Tall open weightsHugging Face; provider says designed for local deploymentExact 50B files, hardware fit, runtime, throughput, quality, routing, operations
Coding Venti weightsSeparate coding-specialized buildDo not confuse with general Venti; verify intended runtime and evaluation

Hosted API checklist

Read the live API documentation before implementation. Capture the exact base URL, authentication method, model identifier, request and response schema, streaming or tool support, timeouts, retries, idempotency, error codes, price, quota, and rate limits.

Also verify prompt and output retention, training policy, region, subprocessors, deletion, acceptable use, commercial terms, incident support, and service continuity. A regional URL does not by itself establish legal entity, residency, or compliance coverage.

Local deployment checklist

Pin the Hugging Face revision and record all file checksums. Inspect configuration, tokenizer, base and specialist files, router metadata, license, dependencies, and model card.

Benchmark the actual machine and workload: supported precision, quantization, accelerator, memory, context, prefill, decoding, concurrent users, tool calls, routing, latency, throughput, power, storage, and recovery. “Designed for local deployment” is a positioning statement, not a universal laptop guarantee.

Use a sandboxed service identity, narrow filesystem and network permissions, secrets management, dependency scanning, logging, budgets, and human review for consequential tool actions.

See the Venti record, Tall record, and Mixture-of-LoRA guide.

Frequently asked questions

How can I access Macaron V1?

Mind Lab lists open weights for Venti, Coding Venti, and Tall on Hugging Face. It also announced separate hosted API entry points for overseas and mainland-China users and describes the API as a way to try Venti without provisioning GPUs.

Which Macaron V1 model is designed for local deployment?

Mind Lab explicitly positions the provider-stated 50B Macaron-V1-Tall for local deployment. Practical fit still depends on the exact files, precision, quantization, runtime, hardware memory, throughput, license, and workload.

Is Macaron V1 available in Flowith?

Mind Lab’s release does not establish a Flowith integration. The provider’s open weights, hosted API, Macaron product surfaces, and Flowith’s current model selector are separate sources and must be checked independently.

Bottom line

Choose by operating responsibility. Hosted Venti reduces provisioning; open weights increase control but transfer serving, security, license, reliability, and cost work to you.

Official source

Source check: August 2, 2026. Verify live endpoints, model IDs, regions, prices, limits, licenses, and Flowith availability.