mirror of
https://github.com/hanzoai/zen.git
synced 2026-11-10 08:21:39 +00:00
3.2 KiB
3.2 KiB
language, license, library_name, pipeline_tag, tags, metrics, model-index
| language | license | library_name | pipeline_tag | tags | metrics | model-index | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
|
apache-2.0 | transformers | text-generation |
|
|
|
Zen 6: Generative Intelligence Model Card
๐ Open Weights ยท Apache 2.0
Developed by Zoo Labs Foundation and served globally on the Hanzo AI Gateway.
Weights:huggingface.co/zenlm
1. Lineup & Model Architecture
Zen 6 is an open-weight foundation model family optimized for agentic coding, deep reasoning, and long-context processing:
| Model ID | Parameters | Architecture | Context Window | Specialization |
|---|---|---|---|---|
zen6 |
70B MoE (8x) | Sparse Mixture of Experts | 131,072 | General reasoning, agentic planning, multimodality |
zen6-flash |
8B Active | Ternary / Quantized Transformer | 65,536 | Ultra-efficient edge & cost-effective serving |
zen6-coder |
32B Dense | Deep Attention Code Encoder-Decoder | 131,072 | Repository-scale code synthesis & automated debugging |
zen-vl |
14B Dense | Multimodal Vision-Language | 32,768 | Visual document analysis, UI inspection, diagram reasoning |
zen-embedding |
430M | 1024-dim Dense Vector Encoder | 8,192 | Semantic search, hybrid RAG, clustering |
zen-reranker |
430M | Cross-Attention Document Reranker | 4,096 | Dense/hybrid RAG reranking and relevance scoring |
2. Verified Benchmark Results
| Benchmark Suite | Metric | Zen 6 | Llama 3.1 70B | Qwen 2.5 72B | DeepSeek-V2.5 |
|---|---|---|---|---|---|
| MMLU-Pro | 5-shot Accuracy | 81.2% | 78.4% | 80.1% | 79.8% |
| HumanEval | Pass@1 (Python) | 89.4% | 82.3% | 86.6% | 85.4% |
| SWE-bench Verified | Resolved Rate | 44.8% | 38.2% | 41.5% | 42.1% |
| MATH | Exact Match | 76.5% | 68.2% | 74.8% | 75.1% |
3. Quickstart with vLLM & Transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "zenlm/zen6"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="bfloat16", device_map="auto")
prompt = "Write an asynchronous event loop in Rust using epoll."
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
4. Hub & Ecosystem Links
- Weights Hub:
huggingface.co/zenlm - Documentation: docs.hanzo.ai/docs/models/zen
- Playground:
huggingface.co/spaces/zenlm/zen-playground