Skip to content
Macaron V1 Venti routes one frozen base to four tiny specialists A diagram showing a single large frozen base model at the bottom, feeding a router in the middle, which picks one of four small 1 billion parameter specialist adapters above it labeled Chat, Agent, Coding, and GenUI for each new request. Macaron V1 Venti One frozen 744B base. Four swappable 1B specialists. MIT · open weights L0 · Chat 1B adapter L1 · Agent 1B adapter L2 · Coding 1B adapter L3 · GenUI 1B adapter L0 router picks one per turn · 0.54s 744B GLM-5.2 base (frozen) shared by all four specialists 753B on the sidebar. 748B on the label. About 775B when you add up the stored tensors.
Guideadvanced

Macaron V1 Venti: a guide to MindLab's 753B mixture-of-LoRA model

The short version

MindLab Research's Macaron V1 Venti pairs one frozen 744B GLM-5.2 base with four 1B specialist adapters. Here is the architecture, the self-reported benchmarks, the real limits, and how to run it.

Published November 21, 2026by
Table of Contents

Macaron V1 Venti: a guide to MindLab's 753B mixture-of-LoRA model

MindLab Research's biggest model has three different sizes, depending on which number on its own Hugging Face page you believe: 748B, 753B, or close to 775B. That is not a vendor rounding up. It is what happens when a small lab ships a genuinely new architecture before the tooling for counting its parameters has caught up.

Macaron V1 Venti pairs one frozen 744B-parameter base model with four tiny, swappable 1B specialist adapters that get routed to per request. We go through what that buys you, what MindLab's own self-reported benchmarks do and do not prove, what it actually costs to run, and whether a lab with no track record on our radar is worth the download.

One frozen base, four featherweight specialists

The architecture is called Mixture of LoRA, or MoL. Instead of one giant model doing everything, Venti sits a 744B-parameter GLM-5.2 base (Z.ai's model family) under four 1B-parameter LoRA adapters for chat, personal-agent tasks, coding, and Generative UI (per MindLab's model card). An L0 router reads each new user turn and picks the specialist to handle it. The base never changes; only a thin adapter swaps in on top.

That three-hop loop has a measured cost. MindLab ran 48 multi-turn mixed-domain requests at temperature 0 and published the breakdown:

HopAvg latencyShare of loop
Route (L0, 24-token constrained decode)0.54s12%
Answer (specialist generation)3.17s68%
Summary (192-token cap, cross-specialist memory)0.97s20%
Total4.68s100%
Source: MindLab Research, Macaron-V1-Venti model card (Hugging Face), fetched November 2026.

Routing hit 99.12% accuracy (6,391 of 6,448 cases) with zero parse errors in that same run, according to MindLab. Worth flagging: that figure comes from the LoRA's own training data, not a held-out test set the model never saw. MindLab calls it "an implementation diagnostic" rather than a generalization estimate, and the residual errors cluster at the boundary between chat and the personal-agent adapter.

The model cannot agree on how big it is

Here is the part that tells you this is week-one tooling for a brand-new shape. The model card's own overview table lists Venti as "748B release label." Summing the tensor shapes in the published adapter files gives roughly 7.69 billion stored values per adapter, putting the four specialists plus the 744B base closer to 774.8B logical parameters. Hugging Face's automatic sidebar counter, meanwhile, reports 753B, which MindLab's own card admits "matches neither number exactly."

Three numbers, one model, and the discrepancy sits in the primary source, not a rumor. It is not disqualifying on its own. It is a sign of how new MoL packaging is: the usual way of counting "how big is this model" assumes one dense checkpoint, and a shared base with swappable adapters does not fit that box cleanly yet.

What MindLab says it can do

Every number below is MindLab's own, pulled from the model card's evaluation table with the asterisks intact; an asterisk marks a score imported from someone else's public leaderboard rather than run under MindLab's own protocol. MindLab Research is a lab we have not covered before with no independent benchmark history on our radar, so read this table as a vendor claim, not a verified ranking.

BenchmarkMacaron V1GLM 5.2GPT 5.5Claude Opus 4.8
ChatBench58.354.555.552.8
SWE-bench Verified85.680.482.9*88.6*
TerminalBench 2.187.682.7*83.4*78.9*
DeepSWE58.454.9*70.0*58.0*
UI4A-Bench87.867.172.175.9
Scores on a 0-100 scale, higher is better. Asterisked values are imported from a public leaderboard or model report as contextual reference rather than run under Macaron's protocol. Source: MindLab Research, Macaron-V1-Venti model card.

That DeepSWE row is worth sitting with: Macaron trails GPT 5.5 by nearly 12 points, and Claude Opus 4.8 beats it on SWE-bench Verified too. MindLab's card says flatly that "metrics and protocols differ across rows" and the table "does not define an aggregate model ranking." The coding and terminal rows also run through a Claude Code agent harness rather than Macaron's own serving stack, so what gets measured is the released system end to end, not the coding adapter alone. Those are the lab's own caveats, not ones we are adding, and they are the right ones to keep.

What it costs, and who can actually run it

The weights are free. Venti ships under the MIT license on Hugging Face, same as the smaller Macaron V1 Tall and the merged Preview checkpoint. Free does not mean cheap to run.

MoL's storage pitch is that it keeps one base resident in memory and loads adapters as thin layers on top, instead of baking each specialist into its own full copy of the base. MindLab puts that at roughly 774.8B logical parameters stored versus 2.976T for four fully merged copies, about 26% of the replicated footprint, according to MindLab. Real efficiency win. It still leaves you serving a 744B-parameter dense-ish MoE base.

MindLab's validated hardware list includes an 8-GPU H20 layout handling 8 concurrent requests at 131K input tokens each, and an 8x B300 cluster with speculative decoding hitting 110 tokens/sec at concurrency 1 and 757 tokens/sec at concurrency 16, according to MindLab. Those are data-center GPU classes, not a workstation card, and the card warns that sharded parallelism layouts "produced systematic corruption" in testing, so a self-hosted deployment has to match the exact engine version and layout MindLab validated.

Skipping that infrastructure, MindLab hosts the model at mintcn.macaron.xin through an OpenAI-compatible endpoint, but the card does not publish pricing there; it tells you to check the site for current rates. Hugging Face's Inference Providers list shows the model is not deployed by any provider yet, and Vultr's catalog routes a Venti deployment request to "Contact Sales" rather than a self-serve hourly rate. For a model this size, what it costs is a number you have to go ask for.

Turning it on

For a quick local test, MindLab's card documents the standard Transformers path:

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("mindlab-research/Macaron-V1-Venti")
model = AutoModelForCausalLM.from_pretrained(
    "mindlab-research/Macaron-V1-Venti", device_map="auto"
)
messages = [{"role": "user", "content": "Who are you?"}]
inputs = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=True,
    return_dict=True, return_tensors="pt",
).to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

That loads the merged weights but skips the routing logic entirely. For real MoL behavior, specialist switching included, MindLab's documented path is its own Mixture of LoRA serving harness, which wraps an OpenAI-compatible endpoint around the L0 router. A plain vLLM server works for the merged checkpoint without the routing:

pip install vllm
vllm serve "mindlab-research/Macaron-V1-Venti"

curl -X POST "http://localhost:8000/v1/chat/completions" \
  -H "Content-Type: application/json" \
  --data '{
    "model": "mindlab-research/Macaron-V1-Venti",
    "messages": [{"role": "user", "content": "What is the capital of France?"}]
  }'

That curl call returns a standard OpenAI-style chat-completions envelope: a choices array with one message.content string and a usage block reporting token counts, the same shape any OpenAI-compatible server returns.

To skip hosting entirely, MindLab's Macaron Artifacts plugin drops the hosted API into an existing agent runtime:

# Claude Code
/plugin marketplace add https://github.com/MindLab-Research/macaron-artifacts
/plugin install macaron@macaron

That gets Generative UI rendering and session viewing against the hosted model without provisioning a GPU, a low-cost way to feel out the routing before committing to self-hosting it.

Where it actually breaks

MindLab's own limitations section is unusually candid, and most of it matters more than the benchmark table. Function calling that requires observing one result before committing the next, the dependent tool-use chains real agents run constantly, scores 49.5% on BFCL v4 versus 54.0% for plain discrete calls, according to MindLab. Long sessions show "qualitative degradation after multiple preference-drift events," flagged as documented but not measured. The routing loop assigns one specialist per turn, so a message mixing a coding question with a scheduling request gets routed to one adapter while the rest waits. There is no standalone safety or red-team evaluation here; MindLab says plainly not to deploy it as an unsupervised decision-maker in medical, legal, or financial settings.

Adoption is its own data point. Venti shows 177 downloads last month and MindLab's Hugging Face organization has 258 followers, modest for a 753-ish-billion-parameter flagship (per Hugging Face, fetched November 2026). Some of that is a size tax: a weight set this large is expensive to even download. But it also means almost nobody outside MindLab has run this in production, and the 11 open discussion threads on the repo are mostly setup questions, not results.

If you want the trick without betting on an unproven lab

MindLab's own family has a smaller option. Macaron V1 Tall runs the same MoL idea, four LoRA specialists routed by an L0 layer, on a Qwen3.6-35B-A3B base instead of a 744B one. At 35B total parameters it fits on hardware a small team can actually provision, with a 262K-token context window.

Skip the routing gimmick entirely and the plain GLM-5.2 base underneath Venti comes from an established lab with its own ecosystem, no MindLab adapter layer required. For a bigger, better-tested open-weight MoE, DeepSeek's V4 Pro, a 1.6T-parameter model, has months more field use and a lab with a long public track record. If what you actually want is LoRA-based task routing on infrastructure you control, Alibaba's Qwen3 family has a mature, widely documented PEFT and LoRA fine-tuning path, so you can build your own routed adapters on a model with a deep community behind it.

The pick

If you are curious what Mixture of LoRA feels like and you have GPU budget to rent by the hour, Macaron V1 Venti is worth a short pilot through the hosted API or the Macaron Artifacts plugin before you spend anything on dedicated hardware. That gets you the routing behavior for the price of API calls, not a GPU cluster.

Shipping a production coding agent this quarter? Skip Venti for now. The coding benchmarks run through a borrowed Claude Code harness rather than Macaron's own stack, the lab has no independent track record, and GPT 5.5 beats it outright on DeepSWE by MindLab's own numbers. DeepSeek V4 Pro or a standard GLM-5.2 or Qwen3 deployment gets you a more proven floor today.

If you are already running GLM-5.2 infrastructure and want to see whether specialist routing earns its keep on your own workload, Macaron V1 Tall is the honest way to test the architecture: smaller, self-hostable, and built on a base you can run without contacting anyone's sales team first.