China Telecom launches a domestic AI model suite deployable on single RTX 3090 GPUs for local enterprise AI inference.

China Telecom may be the first to claim a slice of the AI office automation market.

This summer, major tech companies have collectively doubled down on AI office tools. They're competing over agent capabilities, work entry points, and enterprise workflows—and after coding, office work has become the next battleground for agents.

But while everyone fights for entry points, there's a whole category of enterprises that haven't even stepped through the door.

The reason is painfully practical. Their data cannot leave the corporate perimeter. Models must deploy on-premise. And compute resources are limited. Yet they still need to read long documents, handle complex business processes, and get work done—just like everyone else.

China Telecom's newly open-sourced Xing4.0-29B-A4B targets exactly this gap.

And it's bringing something more: a fully domesticated, lightweight code-intelligent agent LLM. The entire technology stack—from domestic chips and domestic training frameworks, to model training and inference deployment—has been systematically adapted.

To maximize on-premise deployment, Xing4.0-29B-A4B uses a Mixture-of-Experts (MoE) architecture. With 29 billion total parameters, only about 4 billion are activated per inference. After 4-bit quantization, a single RTX 3090 can run it.

But running is just the first step. What enterprises really care about is: what can this model actually do on that graphics card?

For instance, when processing critical documents, the model can complete content understanding, metric extraction, and information organization entirely locally—data never leaves the enterprise environment.

Or take a 200-page procurement bid document. The model can complete initial technical, commercial, and pricing reviews locally in about 20 minutes, with line-by-line scoring justifications.

Going further, local deployment considerations extend beyond consumer-grade GPUs.

Xing4.0-29B-A4B is trained entirely on Huawei Ascend 910C compute, using domestic MindSpore/MindFormers training frameworks and development kits—truly realizing "domestic chips train domestic models."

On the inference deployment side, it has completed Ascend adaptation and verification. From training through production, the domestic compute path is viable.

The model is currently open-sourced across GitHub, Hugging Face, Gitee, ModelScope, and ModelsHub, with concurrent API support.

Notably, as of publication, Xing4.0-29B-A4B has reached #4 on Hugging Face's trending models list.

Interested parties can start using it immediately.

Delivering both performance and capability

Honestly, fitting a model onto one GPU and having it reliably handle daily enterprise work involves plenty more groundwork.

Enterprise tasks are rarely as simple as a single question-answer exchange.

A document gets read, then information must be extracted, tools invoked, results verified, and everything organized into deliverable material.

And this work happens every day—stability and cost-effectiveness matter.

Meeting these requirements means continuing to reduce deployment and operational costs while building genuine capability. The latter is why Xing4.0-29B-A4B focused on strengthening coding and agent abilities.

Start with coding.

Historically, hundred-billion parameter models excel at writing small code snippets or adding a few lines.

But hand over an entire repository and ask it to solve a real engineering problem—things get harder. It needs to understand project structure, trace interface calls across files, locate issues using documentation, logs, and historical context. After fixing the problem, it must verify the changes work and don't break anything else.

Code is just the final output. There's a long chain of work beforehand.

To address this persistent challenge, Xing4.0-29B-A4B advances coding capability from "writing snippets" to "doing engineering."

It natively supports 256K context, expandable to 512K, allowing a single task to fit longer code, documentation, logs, and history—providing space to understand entire projects.

For example, converting scattered contract ledgers from Excel into a governance dashboard requires more than a charting function.

The model must understand the business data in tables, organizing contract summaries, budget distributions, risk identification, and compliance warnings into one page.

This involves data comprehension, code generation, and functional organization—synthesizing information from multiple angles to complete the development.

Xing4.0-29B-A4B can transform these contract ledgers into visual governance dashboards entirely on-premise, with data never leaving the enterprise environment.

Agent capabilities require another level.

After understanding requirements, the model must decompose tasks, sequence execution, invoke tools, and decide next steps based on intermediate results.

Working through the full chain until delivering verifiable results—Xing4.0-29B-A4B received targeted reinforcement for these extended tasks.

For instance, after a user clearly states requirements once, the model can determine which steps to perform first, which can run in parallel, then invoke different tools or agents for weather, calendar, reminders, travel—driving the task forward.

But if the model can handle work itself, can enterprises just plug it in?

One more step: integrating with existing workflows.

The development tools, agent frameworks, and deployment platforms already in use must connect seamlessly. Otherwise, simply switching environments could create mountains of additional work.

To address this, Xing4.0-29B-A4B has completed targeted adaptation for mainstream agent frameworks and harnesses—OpenCode, Claude Code, OpenClaw, Hermes—while remaining compatible with common toolchains like LLaMA-Factory, MindFormers, SGLang, vLLM, and KTransformers, lowering the barrier to integrating with existing development environments.

Of course, for those enterprises mentioned earlier, a real constraint remains unchanged:

Data cannot leave the perimeter, and compute resources are genuinely limited.

So beyond capability upgrades, Xing4.0-29B-A4B continues pushing down deployment barriers.

In traditional FP16 precision, a 29B model consumes roughly 60GB VRAM per GPU, typically requiring multi-GPU setups or expensive A-series cards.

After 4-bit quantization, this shrinks to approximately 15GB—a 75% reduction—allowing consumer-grade RTX 3090 and 4090 to handle it.

This lightweight private deployment solution has already entered real production environments.

In intelligent customer service, Xing4.0-29B-A4B has completed private deployment where user calls, identity information, and business records never leave the premises.

Facing complex requests, the model must complete intent understanding, task decomposition, tool invocation, result synthesis, and compliance verification throughout.

The solution reportedly achieves over 90% success rates for complex business handling.

Here, those enterprises finally have an answer:

Consumer-grade cards can run it. Complex tasks can be completed. Tools and business systems connect seamlessly.

So how does it work?

Comprehensive optimization across architecture, data, training, and deployment

The MoE mentioned earlier is part of the answer.

The model has 29 billion total parameters but activates only 4 billion per inference. This means each task doesn't mobilize all parameters—computational overhead drops accordingly.

But reading long documents, writing engineering code, and sustaining multi-step tasks requires more than MoE.

Architecture, data, training, and engineering optimizations work together.

Consider architecture first.

Longer context means more content fits in. But content occupies space.

When processing long documents, models cache computation results from earlier content for reuse during generation—the familiar KV Cache.

As context lengthens, cache accumulates, consuming VRAM and potentially slowing inference.

For truly usable long context, the cache must shrink.

Xing4.0-29B-A4B uses Multi-head Latent Attention (MLA), which compresses cached information into more compact form, reducing VRAM overhead for long-context inference.

Further, with read overhead down, generation speed must keep pace. Here Xing4.0-29B-A4B employs Multi-Token Prediction (MTP).

Unlike standard training focused on predicting the next token, MTP trains the model to predict multiple subsequent tokens, learning longer dependencies in the process.

These predictions serve inference: generate candidate content in advance, then have the main model verify it, accelerating generation.

Additionally, the model incorporates DeepSeek's manifold hyperbolic connected (mHC) constraint, regularizing information flow across layers to reduce training instability from repeated signal amplification.

With architecture in place, training data matters next.

China Telecom built a data system containing tens of trillions of tokens.

Pre-training data underwent PB-scale multi-source cleaning, incorporating not just general content but complex tables, structured knowledge graphs, and agent behavior trajectories.

The trajectories connect directly to agent capability—they capture exactly how tasks flow from start to finish:

What's the objective? Which tools were called? What did they return? Given results, what's the next action?

Chained together, the model learns how to feed one step's output into the next operation.

In post-training, the team built high-concurrency sandboxes supporting code execution, online search, and tool invocation, constructing extended agent tasks within them.

Whether code runs, tools work correctly, and final results meet requirements—all validated through test cases and automation.

Training material retention now has actual execution results as evidence.

During training, Xing4.0-29B-A4B unfolds with increasing task difficulty.

Pre-training starts at 4K sequence length, building general knowledge and basic reasoning, then gradually raising code and complex reasoning data proportions.

In intermediate training, sequence length extends successively to 32K, 128K, and 256K, adding repository-level code, complex reasoning, and agent data.

In post-training, the team pursued multi-expert reinforcement learning across three directions—Coding, Agent, Reasoning—then fused direction-specific capabilities via MOPD.

Notably, training employed Muon and QK-Clip. The former optimizes parameter updates; the latter controls attention computation magnitude, preventing numerical explosion.

Finally, in the engineering optimization phase, Telecom enabled efficient training on domestic compute.

This is where "fully domesticated" becomes concrete technical detail.

Xing4.0-29B-A4B, based on Ascend 910C clusters with MindSpore/MindFormers, completed mHC feature adaptation, cross-framework precision alignment, and custom operator development, aligning model architecture, domestic training framework, and domestic chips.

For computation scheduling and VRAM pressure, the team applied DVM graph fusion, fine-grained recomputation, and Ascend C custom fusion operators, reducing scheduling and data movement overhead, conserving VRAM, and raising efficiency.

According to the team, after multi-layer coordinated optimization across model architecture, compilation, operators, and hardware, full training throughput improved roughly 96% over baseline—nearly doubled.

The Slime reinforcement learning framework used in post-training also completed Ascend adaptation. From training framework through operators to reinforcement learning, the team applied targeted optimization at every level.

In other words, "domestic chips train domestic models" relies on chips, frameworks, models, and training processes working in concert. And from architecture through data, training, and deployment, this comprehensive effort finally lands on the enterprise tasks stacking up at the desk.

Making AI office work real for enterprises

Finally, to let enterprises truly and seamlessly use AI for office work, Telecom packaged a comprehensive support bundle.

Companies wanting to deploy quickly can pre-integrate Xing4.0-29B-A4B into unified compute-network appliances, reducing environment setup and installation overhead, shortening deployment cycles.

When local compute is tight, enterprises can call cloud compute through dedicated intelligent compute lines, within their data governance requirements, achieving elastic scaling.

For enterprises adopting domestic compute, the model, based on Zhiyuan FlagOS unified open-source AI software stack, has unlocked adaptation paths across multiple domestic chips, lowering migration and deployment costs across hardware platforms.

Different scales, different technical capabilities—enterprises have fitting deployment paths: from a consumer GPU through unified compute appliances to cloud elastic compute.

One final question: why did China Telecom specifically build this lighter model?

Place 29B within the broader Xingchen series, and the answer clarifies.

Previously, China Telecom deployed billion, hundred-billion, and trillion parameter models, exploring trillion-scale parameters alongside speech, semantic, and multimodal directions.

Xing4.0-29B-A4B fills a specific gap:

Limited local compute, yet complex tasks must still be handled.

China Telecom isn't unfamiliar with this need. Honestly, perhaps only Telecom sees and addresses it.

From government services and customer support to network operations, Telecom has long handled extensive private deployment scenarios. Where data can sit, how much computation existing equipment bears, how complex workflows are—these are problems enterprises must solve before deploying models.

This grounded the light model's design.

Starting from enterprise site constraints—compute cost, data boundaries, business requirements—determining model size, identifying essential capabilities, and supplying corresponding delivery options.

This path from model to enterprise matches Telecom's "cloud transformation, digital evolution, intelligent prosperity" strategy and its five-pillar intelligent cloud system: compute, platform, data, models, application.

Models handle understanding and task execution. Platforms manage integration and orchestration. Compute and network enable operation. Industry applications connect these capabilities to specific business.

Xing4.0-29B-A4B is this portfolio's step toward lightweight private deployment.

Trained on domestic chips, then adapted to enterprise equipment and actual workflows, the value of "domestic chips train domestic models" comes down to: can enterprises use it? Will they like it?

And that's why the Xingchen series roadmap looks more tangible: how complex tasks can larger models handle? What work can different modalities cover? How many enterprises previously limited by equipment and deployment constraints can now use lighter models?

Back to the start: AI office competition continues.

Those enterprises where data cannot leave the perimeter, where compute is limited—they keep reading documents, submitting materials, conducting business.

Now they have another way to use AI.

The next step might start with a GPU already on the desk.