Zerre-35B-A3B

Zerre-35B-A3B is an independent, community-developed low-bit derivative of Qwen3.6-35B-A3B. The project explores reducing model storage and memory requirements while preserving as much of the upstream model's text, vision, reasoning, coding, and tool-use capability as possible.

Project status: Work in progress. No model weights, verified benchmark results, or production-ready release are currently published in this repository. This notice will be updated when reproducible artifacts become available.

Zerre is not affiliated with, endorsed by, or sponsored by Alibaba Cloud, the Qwen Team, or PrismML. “Qwen” is used only to identify the upstream model.

Model details

Field	Value
Project	Zerre-35B-A3B
Upstream model	Qwen/Qwen3.6-35B-A3B
Architecture	Multimodal Mixture of Experts (MoE)
Upstream scale	35B total parameters, approximately 3B active parameters per token
Intended optimization	End-to-end low-bit weight compression
Languages	Multilingual; capability follows the upstream model and must be independently evaluated
License	Apache License 2.0
Status	Research and development

The parameter count in the name describes the model architecture, not its file size or memory footprint. Exact weight representation, effective bits per weight, artifact size, supported runtimes, and hardware requirements will be documented separately for each published variant.

Planned release information

Every released variant should include:

• the exact compression or training procedure;
• weight format and effective bits per weight;
• checksum and artifact size;
• supported inference runtime and minimum version;
• memory and throughput measurements with reproducible settings;
• comparisons against the unmodified upstream checkpoint;
• evaluations covering text, vision, reasoning, coding, tool use, safety, and multilingual performance;
• known regressions and unsupported features.

No retention, speed, memory, or quality claim should be inferred until those results are published here.

Usage

Usage instructions will be added after the first verified checkpoint and its inference implementation are published. Standard transformers, llama.cpp, MLX, or other compatibility is not guaranteed unless explicitly listed for a specific artifact.

Do not attempt to load an artifact using upstream Qwen instructions unless the release notes confirm that the artifact uses an upstream-compatible format.

Intended uses

Zerre-35B-A3B is intended for:

• research into low-bit and memory-efficient model deployment;
• local and edge inference experiments;
• evaluation of quality–memory–throughput trade-offs;
• development and testing in controlled environments.

Out-of-scope and high-risk uses

This project is not presented as suitable for decisions that materially affect a person's rights, safety, health, employment, credit, education, housing, or access to essential services. Human review and domain-specific validation are required for high-impact applications. Users are responsible for complying with applicable laws, regulations, platform rules, and third-party rights in their jurisdiction.

Limitations and safety

Compression can introduce regressions that do not appear uniformly across tasks or languages. Compared with the upstream checkpoint, a low-bit variant may show reduced factual accuracy, instruction following, long-context stability, visual understanding, structured output reliability, tool-calling accuracy, or safety behavior.

Like other generative models, Zerre may produce incorrect, biased, unsafe, or fabricated content. Outputs must be independently verified before consequential use. Publishing this model does not constitute a warranty of fitness for any particular purpose.

Evaluation

No verified Zerre evaluation results have been published yet. Future results will identify the exact checkpoint, runtime, prompts, decoding settings, hardware, datasets, and upstream baseline used in each comparison.

Upstream attribution

Zerre-35B-A3B is derived from Qwen3.6-35B-A3B by the Qwen Team at Alibaba Cloud. Please review the upstream model card for its architecture, capabilities, usage guidance, limitations, and citation.

The modifications made by the Zerre project are, or will be, the low-bit conversion/training process, resulting weight files, packaging, documentation, and runtime integration described in each release. The upstream project is not responsible for these modifications.

License

This repository is distributed under the Apache License 2.0. The license permits commercial and non-commercial use, modification, and redistribution subject to its terms, including preservation of applicable license and attribution notices and clear identification of modified files.

The Apache License 2.0 does not grant trademark rights. Names and marks belonging to Alibaba Cloud, Qwen, or other parties remain the property of their respective owners. Datasets, runtime dependencies, and third-party components may have separate terms; users must review them independently.

Citation

If you use the upstream model, cite the Qwen Team as requested in the upstream model card. A project-specific citation will be added with the first versioned Zerre release.

@misc{qwen36_35b_a3b,
  title  = {Qwen3.6-35B-A3B: Agentic Coding Power, Now Open to All},
  author = {Qwen Team},
  year   = {2026},
  month  = {April},
  url    = {https://qwen.ai/blog?id=qwen3.6-35b-a3b}
}

Responsible disclosure

Please use the Hugging Face repository's Community tab to report reproducible model issues. Do not publish private data, credentials, exploit details, or other sensitive information in public reports.