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

Libraries

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

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

pipe = pipeline("text-generation", model="mindlab-research/Macaron-V1-Coding-Venti")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("mindlab-research/Macaron-V1-Coding-Venti")
model = AutoModelForCausalLM.from_pretrained("mindlab-research/Macaron-V1-Coding-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=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use mindlab-research/Macaron-V1-Coding-Venti 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-Coding-Venti"
# 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-Coding-Venti",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

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

SGLang

How to use mindlab-research/Macaron-V1-Coding-Venti 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-Coding-Venti" \
    --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-Coding-Venti",
		"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-Coding-Venti" \
        --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-Coding-Venti",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use mindlab-research/Macaron-V1-Coding-Venti with Docker Model Runner:
```
docker model run hf.co/mindlab-research/Macaron-V1-Coding-Venti
```

Macaron-V1-Coding-Venti

🧭 Parent model: Macaron-V1-Venti
📖 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-Coding-Venti is the coding-specialist checkpoint in the Macaron-V1 family. It is produced by merging the Macaron-V1-Venti L2 Coding LoRA specialist into the GLM-5.2 BF16 base model, yielding a single merged checkpoint for coding, terminal, repository, and software-engineering workflows.

This repository is intended for deployments that want the coding specialist as a standard merged model without runtime LoRA routing. For the full routed Chat, Agent, Coding, and GenUI Mixture of LoRA (MoL) system, use Macaron-V1-Venti.

Highlights

A merged GLM-5.2 coding-specialist checkpoint derived from the Macaron-V1-Venti L2 Coding LoRA.
No runtime adapter loading is required: the LoRA update is merged directly into the BF16 base weights.
Built for code understanding, repository-level software engineering, terminal use, and coding-agent workflows.
Shares the Macaron-V1 post-training stack and evaluation harness used for the flagship Macaron-V1-Venti release.

Model Overview

Field	Value
Model name	Macaron-V1-Coding-Venti
Organization	MindLab Research
Release family	Macaron-V1
Base model	GLM-5.2
Parameter footprint	744B: the L2 Coding LoRA is merged into the GLM-5.2 base; merging adds no parameters
Source specialist	Macaron-V1-Venti L2 Coding LoRA
Checkpoint type	BF16 merged checkpoint
Runtime architecture	GLM-5.2 weights with coding LoRA update merged into the base
Primary domains	Coding, terminal workflows, software-engineering agents
Context length	1M
License	MIT

On parameter counts. The routed Macaron-V1-Venti carries the base plus four separate specialists (744B + 4 × 1B, labeled 748B). Merging one specialist into the base produces a single checkpoint at the base's 744B footprint. The sidebar figure on this page is an automatic safetensors tensor count and may differ from this architectural figure.

Evaluation

This repository is a merged derivative of Macaron-V1-Venti's L2 Coding specialist. It is excluded from the evaluation tables of the technical report (Section 6.1), which reports the routed Macaron-V1-Venti system. The coding and terminal scores below are the parent model's results, provided as context for what the L2 specialist contributes within the routed system; they are not measurements of this merged checkpoint.

Benchmark (Macaron-V1-Venti, routed)	Macaron V1	GLM 5.2	GPT 5.5	Claude Opus 4.8	Gemini 3.1 Pro	Qwen 3.7 Max	Minimax M3
SWE Verified	85.6	80.4	82.9*	88.6*	80.6*	80.4*	80.5*
TerminalBench 2.1	87.6	82.7*	83.4*	78.9*	70.7*	73.5*	66.0*
DeepSWE	58.4	54.9*	70.0*	58.0*	10.0*	18.0*	20.0*
SWE Atlas QnA	49.5	48.9*	45.4*	57.3*	13.5*	22.6	37.9

Higher is better. * marks a value imported from a public leaderboard or model report. Macaron scores use the Claude Code agent scaffold and the retry policies in Appendix B of the technical report, not the production MoL harness. Merging changes the serving path — no router, no cross-specialist summaries, no per-adapter conversation views — so results on this checkpoint may differ from the routed system. Standalone evaluation of this merged checkpoint is not yet available.

Evaluation Protocols

The coding benchmark table is mirrored in evaluation/coding_benchmark_summary.yaml for reproducibility and downstream parsing.

The full benchmark methodology is documented in Appendix B of the technical report.

Limitations

Not evaluated as a standalone checkpoint. The technical report excludes this merged build from all evaluation tables and claims. Reported Macaron-V1 coding results describe the routed Venti system.
Coding only. This checkpoint carries the L2 Coding specialist. It has no Chat, Agent, or GenUI specialist, no L0 router, and no cross-specialist summary mechanism. For personal-agent workflows, conversational use, or UI4A Generative UI, use Macaron-V1-Venti with the Mixture of LoRA serving harness.
No continual-learning affordances. A merged checkpoint gives up the two properties that motivate MoL: base-weight immutability under further specialization, and the ability to add or update a specialist by adapter registration rather than base retrain.
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 remain a known weak spot.
No component-level attribution. The parent model's coding rows evaluate the released system end to end and do not isolate the contribution of the L2 adapter, so they do not predict this checkpoint's behavior component by component.

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. This checkpoint has no router and no NoAI visibility boundary enforcement from the UI4A action contract, since those live in the harness rather than the weights. If you rely on those boundaries, serve the routed system instead. Inherited requirements from the GLM-5.2 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

Deploy as a standard GLM-5.2-compatible BF16 merged checkpoint; no PEFT adapter attachment or MoL router is required. Sizing follows the 744B base. The validated long-context envelope measured for the routed configuration also applies to the base weights: on H20 with TP4/PP2/DCP4, 16 concurrent 56K-token requests, 8 × 180K, or 4 × 230K completed cleanly; on B300 TP8, use replicated DCP layouts, since sharded DCP produced systematic corruption. FlashMLA sparse attention produced clean outputs across all 48 tested long-context configurations, whereas the default DSA decode path was clean in only 6 of 48 on GLM-5.1. Because this build has no adapters resident, it does not benefit from MoL's base-sharing residency advantage: four merged specialists would require four full base copies (2.976T parameters) versus ~774.8B for the routed layout.

Usage

This repository contains a merged checkpoint. Load it as a standard GLM-5.2-compatible causal language model; no PEFT adapter attachment or MoL router is required.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "mindlab-research/Macaron-V1-Coding-Venti"

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()

For production serving, use the same GLM-5.2-compatible serving stack you would use for a merged BF16 checkpoint. The full Macaron routed system, including Chat, Agent, Coding, and GenUI specialists, remains available through Macaron-V1-Venti and the Mixture of LoRA (MoL) serving harness.

Macaron Artifacts

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

Install it as a plugin in the agent runtime you use:

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

# Codex
codex plugin marketplace add https://github.com/MindLab-Research/macaron-artifacts
codex plugin add macaron@macaron

# Kimi Code
/plugins install https://github.com/MindLab-Research/macaron-artifacts
/reload

License

This repository is released under the MIT License. Users should also respect any requirements inherited from the GLM-5.2 base model and from dependencies used by their serving stack.

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}
}