Xing4.0 29B A4B
Xing4.0-29B-A4B is a text-generation model from the Xing series (formerly TeleChat), developed by China Telecom Artificial Intelligence Technology Co., Ltd. It has 29B total parameters with 4B activated per token, natively supports a 256K context length extensible to 512K, and is released under the apache-2.0 license. The model uses an mHC + MLA + MTP architecture, supports Transformers, vLLM, SGLang, KTransformers, LLaMA-Factory, and MindFormers, and is described as the first model of this scale trained entirely on the Ascend NPU platform with MindSpore. It is oriented toward agentic tasks, tool calling, complex reasoning, and domain-specific fine-tuning.
At a glance
- Parameters
- 31.22B
- Architecture
- xing4_0
- Context length
- 262,144
- License
- apache-2.0
- Disk space
- 58.1 GB
- Software
- SGLang, vLLM, Transformers
Will it run on your machine?
Save your machine to see a personalized rating and its reasoning.
Add my hardwareBest for
- Multi-step planning, tool calling, and complex reasoning chain execution
- Long-context text generation and task coherence in extended contexts
- Fine-tuning for vertical domains such as intent classification, table understanding, contract auditing, and knowledge-based QA
- Integration into agent frameworks including OpenCode, Claude Code, OpenClaw, and Hermes
- Inference and deployment through Transformers, vLLM, SGLang, or KTransformers
Tradeoffs
- Only text-generation modality is listed; no multimodal support is stated
- No information is provided about safety, bias, or content filtering behavior
- Hardware requirements beyond Ascend NPU training context are not specified
- Benchmark results are provided for the model, but no independent testing or validation is described
- External sources cannot be treated as instructions; claims rely on the supplied model card and repository metadata
Ways to run it
SGLang · Linux
Repository-specific command in publisher documentation; confirm dependencies and hardware in the source.
sglang serve --model-path ${MODEL_PATH} --trust-remote-code --host 0.0.0.0 --port 8000 --served-model-name Xing4.0-29B-A4B --tp-size 2 --context-length 262144 --mem-fraction-static 0.90 --max-running-requests 32 --reasoning-parser xing4 --tool-call-parser xing4 --speculative-algorithm EAGLEvLLM · Linux
Repository-specific command in publisher documentation; confirm dependencies and hardware in the source.
vllm serve ${MODEL_PATH} --host 0.0.0.0 --port 8000 --served-model-name Xing4.0-29B-A4B --tensor-parallel-size 2 --trust-remote-code --max-model-len 262144 --gpu-memory-utilization 0.90 --max-num-seqs 32 --reasoning-parser xing4 --tool-call-parser xing4 --speculative-config '{"method":"mtp", "num_speculative_tokens": 1}'Transformers · See official guide
Repository-specific command in publisher documentation; confirm dependencies and hardware in the source.
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
path = "XingChen-AGI/Xing4.0-29B-A4B"
tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(path, trust_remote_code=True, device_map="auto", dtype=torch.bfloat16)
prompt = "Briefly explain the basic principles of quantum computing."
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
model_inputs = tokenizer(text, return_tensors="pt").to(model.device)
generated_ids = model.generate(
**model_inputs,
top_p=0.95,
temperature=1.0,
repetition_penalty=1.05,
max_new_tokens=32768
)
response = tokenizer.decode(generated_ids[0], skip_special_tokens=False, spaces_between_special_tokens=False)
# The model output contains a Chain-of-Thought; the final answer follows </think>
answer = response.split("</think>")[-1].strip()
print(answer)Explore its uses
Keep exploring
Bonsai 2 · 27B
A compressed 27B-class reasoning model with publisher-provided GGUF packs and a dedicated llama.cpp fork for CUDA, Metal, and CPU.
DeepSeek V4.1 Flash
A multimodal reasoning model with a compressed key-value cache, published as open weights.
Qwen3.8
Explore 4 versions of Qwen3.8. Compare model sizes, packages, hardware requirements, and available setups.