GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
NapYang
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
NapYang
Model Tree
Input Modalities
Output Modalities
Supported Functionality
Sage-A1.5-UD-4bit-fp16 inherits the full capability stack of Sage-A1.5 — agentic orchestration, verification mode, tool use, and open scientific discovery reasoning — while consuming a fraction of the memory. It is designed for long-horizon task execution with sub-agent delegation, evidence validation, and exploratory research across domains.
mode: affine, group_size: 64)switch_mlp (gating), plus MTP layers' gate/down/up — the largest memory savings come from the MoE switch FFN.lm_head, embed_tokens, all linear_attn projections, and all shared_expert MLP weights.model-mtp.safetensors), vs ~67 GB full precision.| Parameter | Value |
|---|---|
| Architecture | Qwen3_5MoeForConditionalGeneration |
| Layer type | MoE Transformer with switch FFN + linear attention hybrid |
| Hidden size | 2048 |
| Number of layers | 40 (linear attention / full attention: 3/1 per block) |
| Attention heads | 16 (GQA, 2 KV heads) |
| Head dim | 256 (16 linear key heads @ 128 + 32 linear value heads @ 128) |
| Vocabulary size | 248,320 |
| Max sequence length | 262,144 (256K) |
| RoPE theta | 10,000,000 (partial rotary factor 0.25) |
| RoPE layout | mRoPE interleaved, sections [11, 11, 10] |
| MoE experts | 256 total, 8 active per token |
| Expert FFN hidden size | 512 (switch MLP) / 512 (shared expert) |
| Vision encoder hidden size | 1152 → 2048 (out) |
| Norm | RMSNorm, eps 1e-6 |
| Init std | 0.02 |
chat_template_kwargs: {"verification": true} and a non-empty system message).The base model Sage-A1.5 (unquantized) achieves:
See the Sage-A1.5 README for the full evaluation table (GAIA, BrowseComp, DeepResearch-Bench2, SWE-bench Pro/Multilingual, HiPhO, CritPt, ALE-CLI, MatTools).
Load with Transformers (4.57+) in safetensors format:
python
from transformers import AutoModelForCausalLM, AutoTokenizer model = AutoModelForCausalLM.from_pretrained( "NapYang/Sage-A1.5-UD-4bit-fp16", # your repo name device_map="auto", torch_dtype=torch.float16,)tokenizer = AutoTokenizer.from_pretrained("NapYang/Sage-A1.5-UD-4bit-fp16")Recommended sampling parameters (from base model card):
temperature: 1.0top_p: 0.95top_k: 20min_p: 0.0presence_penalty: 1.5repetition_penalty: 1.0Pass a non-empty system message plus chat_template_kwargs: {"verification": true}; the model pauses mid-trajectory and reviews its own chain of thought. Read reasoning_content in addition to content.
python
message = model.chat( tokenizer, messages=[ {"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "Find the population of Beijing and cross-check it."}, ], tools=TOOLS, chat_template_kwargs={"verification": True},)print(message.reasoning_content) # self-review happens inside the chain of thoughtprint(message.content)PreTrainedProcessor) from the base repo, not just the tokenizer.chat_template.jinja); make sure to load it alongside this checkpoint.If you find our work helpful, feel free to give us a cite.
text
@misc{lou2026sagea15longhorizonexecutionscientific, title={Sage-A1.5: From Long-Horizon Execution to Scientific Exploration with Capability Composition}, author={Sage-A1.5 Team}, year={2026}, eprint={2610.28578}, archivePrefix={arXiv}, primaryClass={cs.AI}, url={https://arxiv.org/abs/2610.28578},}This model is a derivative of InternScience/Sage-A1.5 (Shanghai AI Laboratory, Apache 2.0). Base model built on Qwen3.6 (Alibaba Cloud, Apache 2.0) and Agents-A1.
from transformers import AutoModelForCausalLM, AutoTokenizer model = AutoModelForCausalLM.from_pretrained( "NapYang/Sage-A1.5-UD-4bit-fp16", # your repo name device_map="auto", torch_dtype=torch.float16,)tokenizer = AutoTokenizer.from_pretrained("NapYang/Sage-A1.5-UD-4bit-fp16")message = model.chat( tokenizer, messages=[ {"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "Find the population of Beijing and cross-check it."}, ], tools=TOOLS, chat_template_kwargs={"verification": True},)print(message.reasoning_content) # self-review happens inside the chain of thoughtprint(message.content)@misc{lou2026sagea15longhorizonexecutionscientific, title={Sage-A1.5: From Long-Horizon Execution to Scientific Exploration with Capability Composition}, author={Sage-A1.5 Team}, year={2026}, eprint={2610.28578}, archivePrefix={arXiv}, primaryClass={cs.AI}, url={https://arxiv.org/abs/2610.28578},}