Hosted Venti API
Evaluate the flagship without owning a 748B serving stack.
Verify API documentation, model name, region, authentication, pricing, rate limits, retention, and service terms live.
Mind Lab · flagship model record Source check · August 2, 2026
Quick answer
Macaron V1 Venti is open-weight, but not lightweight.
Mind Lab describes Venti as a 748B flagship built from a 744B GLM-5.2 base and four 1B LoRA specialists. Use the hosted API for bounded evaluation or plan a serious serving program for the open weights; do not infer cheap local deployment from availability alone.
Verified model record
Provider
Mind Lab; Macaron-V1-Venti was announced July 21, 2026.
Provider-stated size
748B parameters: a 744B GLM-5.2 base plus four 1B LoRA specialists.
Specialists
L0 Chat, L1 Agent, L2 Coding, and L3 GenUI, selected through a routing layer.
Weights
Mind Lab lists Venti in its open-weight Macaron-V1 Hugging Face collection.
Hosted access
Mind Lab provides separate hosted API entry points for overseas and mainland-China users and describes the service as a way to try Venti without provisioning GPUs.
Flowith access
Not established by Mind Lab's release; verify the current Flowith model selector separately.
Choose by operating model
Hosted evaluation or infrastructure ownership
Self-hosted open weights
Teams with the infrastructure and controls to inspect or operate the released weights.
Open weights do not make a 748B model inexpensive or turnkey; validate license, hardware, serving, quantization, security, and support.
Evaluation gate
Open weights are the start of verification
Mind Lab publishes architecture and evaluation detail, but production selection still needs independent workload testing and an explicit operating model.
- 01Treat parameter counts, base-model identity, benchmark results, and self-hostability as provider claims until independently reproduced.
- 02Do not turn LongStraw's multi-million-token training result into an undocumented Venti inference context-window claim.
- 03Evaluate routing quality separately across chat, tool use, coding, and generative UI; a specialist architecture does not guarantee the correct route.
- 04Test tool-call correctness, long-horizon recovery, code execution, UI output, latency, throughput, and total serving or API cost on representative tasks.
- 05Review the current weight license, dependencies, base-model obligations, data policy, acceptable use, and commercial terms before deployment.
- 06Verify hosted API access, local serving, the Macaron consumer experience, and Flowith availability as separate product surfaces.
Continue by decision intent
Official source
Mind Lab: Introducing Macaron-V1 — variant identities, provider-stated parameters, architecture, evaluation, open weights, hosted API, and access paths.
Macaron V1 Venti questions, answered
Macaron-V1-Venti is Mind Lab's provider-described 748B flagship agent model, built from a 744B GLM-5.2 base and four 1B LoRA specialists for chat, agents, coding, and generative UI.
Mind Lab lists Venti in the Macaron-V1 open-weight collection on Hugging Face. Check the current repository, license, files, checksums, dependencies, and usage conditions before downloading or deploying it.
Mind Lab announced hosted API entry points for overseas and mainland-China users and specifically describes them as the fastest way to try Venti without provisioning GPUs. Verify the live docs for model IDs, prices, limits, and availability.
Open weights allow self-hosting in principle, but a provider-stated 748B model requires substantial infrastructure. Measure hardware, memory, quantization, throughput, reliability, and operating cost before calling it practical.
Mind Lab's release does not establish a Flowith integration. Provider weights, hosted API access, the Macaron product, and Flowith's current model selector are separate sources.