01
Overview
Xing4.0-29B-A4B is a lightweight agent-oriented large language model from China Telecom Artificial Intelligence Technology. The model was open-sourced on September 17, 2026, with standard, FP8 and GGUF artifacts listed by the official project.
02
What is Xing4.0-29B-A4B?
Xing4.0 is the latest generation of the TeleChat line. The official repository describes the 29B-A4B model as a mixture-of-experts system with 29 billion total parameters and about 4 billion parameters activated per token. Its architecture combines mHC, MLA and MTP, and the project is explicitly optimized for agent workloads such as multi-step planning, tool calling and long-running execution.
The model natively supports a 256K context window, which the developers say can be extended to 512K. That makes long-context coding, document analysis and agent workflows central parts of the release rather than secondary features.
03
Open model formats and ecosystem
China Telecom AI released the base model alongside official FP8 and GGUF variants. The project documents inference/deployment paths for vLLM, SGLang and KTransformers, and fine-tuning support for LLaMA-Factory, MindFormers and FlagOS. It also says it has performed targeted format alignment for agent environments including OpenCode, Claude Code, OpenClaw and Hermes.
This breadth matters because the practical value of an open model depends not only on weights but also on whether developers can deploy, quantize, fine-tune and connect it to the tools they already use.
04
Can Xing4.0 run locally?
Yes, local deployment is an explicit part of the release, and an official GGUF is available. However, there is no single honest RAM or VRAM requirement for every setup. Memory consumption changes with the selected artifact and quantization, context length and KV cache, runtime, batch size and concurrency.
Launch material says 4-bit quantization can reduce model memory to roughly 15 GB and frames 24 GB consumer GPUs such as the RTX 3090 and RTX 4090 as viable targets. Treat that as a release/vendor claim, not a guarantee that every 256K- or 512K-context workload will fit comfortably on a 24 GB GPU. Long context can add substantial KV-cache memory.
A separate local-running guide should handle exact runtime commands, quantization choices and cautious hardware planning.
05
Context length
The official repository states:
- native context: 256K tokens
- extensible context: 512K tokens
An advertised maximum context window should not be confused with a recommended context for every machine. Developers deploying locally should choose context based on available memory and latency requirements.
06
Agent and coding capabilities
The model is designed around execution rather than simple chat. The project highlights complex task understanding, planning, tool calls, multi-step reasoning and stable execution across long contexts. Potential uses include coding agents, research, data analysis, document-heavy automation and private enterprise workflows.
07
Benchmarks
China Telecom AI reports 75.0 on SWE-bench Verified, 66.0 on SWE-bench Multilingual and 57.5 on Terminal-Bench 2.1. It also reports 76.55 on Claw-Eval, 60.80 on DeepresearchBII and 64.63 on Tau3-Bench.
These numbers come from the model's own published evaluation and should be read as vendor-reported results until independently reproduced under comparable settings. The repository compares Xing4.0 with Qwen3.6-35B-A3B and Gemma4-26B-A4B, but benchmark tables alone are not enough to declare a universal winner.
08
Training stack
The developers say Xing4.0 was trained entirely on Huawei Ascend NPU infrastructure using MindSpore. They describe it as the first model at this scale trained entirely on the Ascend NPU platform with MindSpore and optimized for complex engineering tasks. Claims about training-throughput improvements are engineering results reported by the project and are not independent benchmarks.
09
Who is Xing4.0 for?
Xing4.0 is especially interesting for developers who want an agent-focused open model with low active parameter count, long context, official GGUF availability and support across multiple inference stacks. It may also appeal to organizations evaluating private/local agent deployments where data should remain inside their own infrastructure.
10
Important limitations
The 512K figure is an extension beyond the native 256K context and should be described that way. Consumer-GPU feasibility depends strongly on quantization and context. Benchmark results are developer-reported. Runtime support can change quickly after a model launch, so commands and compatible versions should be rechecked before deployment.
11
FAQ
Is Xing4.0-29B-A4B open source?
The model weights and project artifacts are publicly available through the official project and model hubs. Recheck the exact repository/model license before commercial deployment because license terms can change independently of technical availability.
Is there an official Xing4.0 GGUF?
Yes. The project lists an official Xing4.0-29B-A4B-GGUF release.
How many parameters does it use?
The model has 29B total parameters with roughly 4B activated per token.
What is the context window?
The documented native context is 256K tokens, extensible to 512K.
Does it support vLLM and SGLang?
The official project documents both vLLM and SGLang, as well as KTransformers.
Can it work with coding-agent tools?
The developers document targeted adaptation/format alignment for OpenCode, Claude Code, OpenClaw and Hermes.
12
Sources
Primary technical details and benchmark numbers come from the official XingChen-AGI repository and official Hugging Face model artifacts. Launch-date and consumer-GPU claims are cross-checked against the China Telecom AI launch release. Benchmark and performance claims remain attributed to their publishers.
Sources
Primary and supporting sources
Use the linked documentation to check current pricing, access and product behaviour. The dates above describe this article; they do not guarantee that every vendor claim remains current.