Instructions to use mindlab-research/Macaron-V1-Tall with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use mindlab-research/Macaron-V1-Tall with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="mindlab-research/Macaron-V1-Tall")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("mindlab-research/Macaron-V1-Tall")
model = AutoModelForMultimodalLM.from_pretrained("mindlab-research/Macaron-V1-Tall", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.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=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use mindlab-research/Macaron-V1-Tall with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "mindlab-research/Macaron-V1-Tall"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "mindlab-research/Macaron-V1-Tall",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/mindlab-research/Macaron-V1-Tall

SGLang

How to use mindlab-research/Macaron-V1-Tall with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "mindlab-research/Macaron-V1-Tall" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "mindlab-research/Macaron-V1-Tall",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "mindlab-research/Macaron-V1-Tall" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "mindlab-research/Macaron-V1-Tall",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use mindlab-research/Macaron-V1-Tall with Docker Model Runner:
```
docker model run hf.co/mindlab-research/Macaron-V1-Tall
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Macaron-V1-Tall

📖 Blog: Introducing Macaron-V1
🚀 Hosted API: Macaron API Platform
🧩 Artifacts: Macaron Artifacts
🛠️ Self-hosted serving: Mixture of LoRA (MoL) serving harness
📄 Technical report: Macaron-V1: Towards Open Continual Learning with Self-Improvement and Mixture-of-LoRA (arXiv:2608.09819)
✉️ Correspondence: contact@mindlab.ltd

Macaron-V1-Tall is a Mixture of LoRA (MoL) model in the Macaron-V1 family, built on Qwen3.6-35B-A3B for personal intelligence, tool use, coding workflows, and code-native Generative UI.

The model combines a Qwen3.6 MoE base with the Macaron-V1 specialist system: four LoRA specialists for chat, personal-agent tasks, coding, and GenUI, with an L0 router selecting the most suitable specialist for each new user request.

Macaron-V1-Tall is released alongside Macaron-V1-Venti as a sibling model in the Macaron-V1 family. Both models share the same product-facing specialist layout and routed serving interface while targeting different deployment and experimentation profiles.

Highlights

A Qwen3.6-35B-A3B-based Macaron-V1 model with the full Macaron specialist system.
Mixture of LoRA (MoL) architecture with four specialists across Chat, Agent, Coding, and GenUI.
Built for personal-agent workflows, tool use, repository-level coding, and UI4A Generative UI.
BF16 checkpoint with four LoRA adapters stored under loras/L0 through loras/L3.
Supports a 262K context length according to the released checkpoint configuration.

Model Overview

Field	Value
Model name	Macaron-V1-Tall
Organization	MindLab Research
Release family	Macaron-V1
Base model	Qwen3.6-35B-A3B
Architecture	Qwen3.6 MoE base + Mixture of LoRA (MoL) specialists
Parameter footprint	50B release label: a nominal 35B Qwen3.6-35B-A3B base plus four LoRA specialists (~50.1B by nominal-base-plus-adapter count)
LoRA rank / alpha	r = 64, α = 128, dropout 0
Stored values per adapter	3,775,651,840 (BF16 for L0/L1/L3; F32 for L2)
Specialists	L0 Chat, L1 Agent, L2 Coding, L3 GenUI
Post-training system	MinT + MindForge
Primary domains	Personal intelligence, tool use, coding, Generative UI
Context length	262K
Text config	40 layers, 2048 hidden size, 16 attention heads, 2 KV heads, 256 experts, 8 experts per token
Precision / serving format	BF16 base checkpoint with routed LoRA serving
License	MIT

On parameter counts. Tall's adapter configurations use a different rank and expert target set from Macaron-V1-Venti, so its aggregate is reported as a rounded release footprint rather than inferred from base size alone. The two adapter budgets are not directly comparable because the bases expose different target-module and expert structures. The sidebar figure on this page is an automatic safetensors tensor count and may differ from this architectural figure.

Mixture of LoRA (MoL) Architecture

Adapter	Role	Description
`loras/L0`	Chat	Conversational and instruction-following backbone; entry point for routing.
`loras/L1`	Agent	Personal-life agent tasks, heavy tool use, long-horizon planning, and dynamic workflows.
`loras/L2`	Coding	Code understanding, SWE tasks, terminal use, and repository workflows.
`loras/L3`	GenUI	UI4A rendering and UI-driven action.

At runtime, L0 routes each new user request to the most suitable specialist. Ongoing reasoning and tool interactions remain within the selected LoRA, while completed work can be shared across specialists through concise summaries.

Routing Behavior and Cost

Hop	Avg latency	Share of loop
Route (L0 constrained decode, 24 tok)	0.20 s	11%
Answer (specialist generation)	1.24 s	70%
Summary (192-tok cap)	0.32 s	19%
Total	1.76 s	100%

Measured over 48 multi-turn mixed-domain requests at temperature 0. The routing overhead share (~30%) is stable across base sizes, while absolute loop latency is roughly 2.7× lower than Macaron-V1-Venti's 4.68 s.

On the same 6,448-sample trace used for Venti — verified identical by matching every sample ID and input hash — Tall reaches 99.04% routing accuracy with 100% canonical-label compliance, so routing quality is stable across base sizes. L3 GenUI reaches 100% and L1 Agent is lowest at 95.1%. The trace is drawn from LoRA training data, so this is an implementation diagnostic rather than a generalization estimate.

Evaluation

Macaron-V1-Tall compared with its Qwen3.6-35B-A3B base on the seven benchmarks evaluated for both systems. Both use the same protocol within each row; higher is better, normalized to 0–100.

Benchmark	Macaron-V1-Tall	Qwen3.6-35B-A3B
Macaron ChatBench	54.9	48.0
Macaron LivingBench	48.4	47.1
PinchBench	86.2	82.5
ClawGym	64.0	58.6
SWE-bench Verified	75.4	73.4
TerminalBench 2.1	56.2	52.5
UI4A-Bench (Final Score)	59.3	33.9

Tall has a larger point estimate than its base on all seven rows, from +1.3 on Macaron LivingBench to +25.4 on UI4A-Bench. This is an end-to-end system comparison, not a parameter-matched component ablation: the two systems differ in parameter footprint, adapters, routing, and potentially harness behavior. The table does not identify which component produces each difference, and it does not establish scaling behavior from Macaron-V1-Tall to Macaron-V1-Venti.

The headline Macaron-V1 comparison against frontier baselines is reported for Macaron-V1-Venti; see the Venti model card and Table 8 of the technical report. Per-benchmark judges, case counts, and metrics are documented in Appendix B of the report.

Multimodal Behavior

The Macaron-V1-Tall specialists are trained on text-only data, while the Qwen3.6-35B-A3B base is multimodal. The following full-dataset point estimates (no-thinking mode) compare the routed MoL service with the native base:

Benchmark	Base	Tall (MoL)	Δ
OCRBench v1 (%)	88.80	89.60	+0.80
MMBench-EN dev	86.08	86.68	+0.60
MMMU val (%)	59.89	61.22	+1.33
MME perception	1785.56	1732.57	−52.99
MME cognition	604.64	671.07	+66.43

These measurements do not include repeat-level variance, vision-based agent tasks, or an adapter/routing ablation. They therefore do not establish preservation of multimodal capability, cross-modal transfer, or a mechanism for the observed differences. Vision inputs are not covered by the specialist training data, so treat multimodal behavior as inherited base capability rather than a tuned capability of this release.

Limitations

Base-vs-system comparison only. The reported evaluation compares the deployed Macaron-V1-Tall system with its base configuration. It is not a parameter-matched ablation and does not establish scaling behavior between Tall and Macaron-V1-Venti.
Multimodal capability is inherited, not tuned. The specialists are text-only; see Multimodal Behavior above for the measurement boundary.
One snapshot, not a demonstrated learning curve. This release is a single snapshot of the recursive self-improvement loop. It cannot by itself distinguish compounding improvement across generations from a single round of self-generated-data training; cross-generation lift is not measured.
No collective-intelligence evidence. MoL is designed so that specialists trained by different teams or personalized for different users can compose on one shared base. This release evaluates only the four shipped specialists and does not test a cross-owner adapter population, nor establish robust switching under broad workloads.
Single intent per turn. The routing loop assigns a whole user turn to one specialist. Messages that pack several intents are handled by routing to one specialist and letting the conversation segment over later turns; multi-intent decomposition is exploratory and not in production.
Stateful observe-before-commit APIs. On BFCL v4 (200 tasks), our REPL substrate scores 49.5% versus 54.0% for discrete function calling. Dependent calls that must observe one result before committing the next are a known weak spot; the harness therefore permits falling back to discrete calls or shell.
Long-session character stability. We observe qualitative degradation after multiple preference-drift events in very long sessions. This failure mode is documented but not quantitatively measured.
No component-level attribution. Whether reported gains come from specialization, routing, the harness, or their interaction is not resolved by any controlled experiment in this release.
Reproducibility boundary. A reported run is reconstructable only when the adapter revision and compatible base are joined with the corresponding harness configuration, task-bank version, evaluator, and sampling settings. This card does not carry a complete per-specialist training specification.

Safety

Scope of evaluation. This release does not include a standalone safety or red-team evaluation. Anti-sycophancy is enforced as an independent scored dimension in Macaron ChatBench, and human review is required for harness changes that touch tool exposure or safety-relevant boundaries, but neither substitutes for a dedicated safety assessment. We interpret the reported results as a systems characterization, not as evidence that this release is suitable for safety-critical use.

Data governance. Internal evaluations include de-identified product conversations and traffic. This release does not document the consent or opt-in basis for research use, the de-identification procedure and residual re-identification audit, retention and access controls, or an ethics-review determination.

Deployment guidance. Do not deploy this model as an unsupervised decision-maker in medical, legal, financial, or other high-stakes personal-assistance settings. The UI4A action contract includes a NoAI visibility boundary for fields the model must not see; use it for sensitive data rather than relying on prompt-level instructions. Inherited requirements from the Qwen3.6-35B-A3B base model and from serving-harness dependencies also apply.

Reporting. Please open a discussion on this repository or contact contact@mindlab.ltd for safety issues.

Hardware Requirements

Macaron-V1-Tall targets local deployment and lower-latency serving. The BF16 base checkpoint plus four LoRA adapters comes to a ~50B release footprint, and the routed loop measures 1.76 s end to end versus 4.68 s for the 748B Venti configuration under the same 48-request profile. Serve the base with a Transformers version that supports the Qwen3.6 / Qwen3.5 MoE architecture; for routed serving, use the Mixture of LoRA harness, which keeps the endpoint OpenAI-compatible while adding the L0 router and same-request specialist switching. Quantizations contributed by the community are linked in the model tree on this page.

Training Details

Field	Value
Base	Qwen3.6-35B-A3B
LoRA rank / alpha / dropout	r = 64, α = 128, dropout 0
Target scope	Attention/recurrent and MLP projections (`in_proj_*`, `q_proj`, `k_proj`, `v_proj`, `out_proj`, plus the Venti set), and expert parameters `experts.gate_up_proj`, `experts.down_proj`; excludes `gate`, `lm_head`, `shared_expert_gate`
Extra saved modules	none (`modules_to_save=null`)
Stored values per adapter	3,775,651,840 (BF16 for L0/L1/L3; F32 for L2)
Optimizer	AdamW, lr 5e-6, batch size 4, linear-warmup cosine schedule, warmup ratio 0.1
Epochs	4 (L0, L2); 1 (L1); 1 with batch size 2 (L3)

Usage

Hosted API

The hosted API is available at https://mintcn.macaron.xin/. Use the site for current model names, authentication, pricing, and rate-limit details.

For OpenAI-compatible deployments, requests follow the standard chat-completions shape:

curl https://mintcn.macaron.xin/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <api-key>" \
  -d '{
    "model": "Macaron-V1-Tall",
    "messages": [
      {"role": "user", "content": "Create a compact dashboard UI for tracking weekly fitness goals."}
    ],
    "temperature": 0.2,
    "max_tokens": 2048
  }'

Transformers

This repository contains the base checkpoint at the repository root and the Macaron LoRA specialists under loras/. Load the base checkpoint with a Transformers version that supports the Qwen3.6 / Qwen3.5 MoE architecture.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "mindlab-research/Macaron-V1-Tall"

tokenizer = AutoTokenizer.from_pretrained(
    repo_id,
    trust_remote_code=True,
)

model = AutoModelForCausalLM.from_pretrained(
    repo_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
model.eval()

Mixture of LoRA (MoL) Serving

For self-hosted routed serving, use the Mixture of LoRA (MoL) serving harness. The harness keeps the endpoint OpenAI-compatible while adding an L0 router, server-side LoRA metadata, and same-request switching into the selected specialist.

See the MoL serving repository for more details.

Macaron Artifacts

Macaron Artifacts is the companion local WebUI and plugin bundle for viewing Macaron sessions and GenUI output. It supports Claude Code, Codex, and Kimi Code, and can run against Macaron or another Anthropic-compatible endpoint.

The plugin includes the genui-builder skill so supported agents can produce GenUI TSX and preview the rendered artifact in the browser. See the Artifacts repository for full install, update, and provider configuration details.

License

This repository is released under the MIT License. Users should also respect any requirements inherited from the Qwen3.6-35B-A3B base model and from dependencies used by the serving harness.

Citation

@article{mindlab2026macaronv1,
  title   = {Macaron-V1: Towards Open Continual Learning with Self-Improvement and Mixture-of-LoRA},
  author  = {{Mind Lab}},
  journal = {arXiv preprint arXiv:2608.09819},
  year    = {2026},
  url     = {https://arxiv.org/abs/2608.09819}
}