{"id":12081,"date":"2026-11-17T11:01:05","date_gmt":"2026-11-17T02:01:05","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/?p=12081"},"modified":"2026-11-17T11:10:11","modified_gmt":"2026-11-17T02:10:11","slug":"xing4-0-29b-a4b-w8a8-released","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/xing4-0-29b-a4b-w8a8-released\/","title":{"rendered":"Xing4.0-29B-A4B-w8a8 Released: W8A8 Quantized MoE Model"},"content":{"rendered":"<p><!-- lmw:facts --><\/p>\n<h2>At a Glance<\/h2>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Item<\/th>\n<th>Value<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Repository<\/td>\n<td><a href=\"https:\/\/huggingface.co\/XingChen-AGI\/Xing4.0-29B-A4B-w8a8\">XingChen-AGI\/Xing4.0-29B-A4B-w8a8<\/a><\/td>\n<\/tr>\n<tr>\n<td>Publisher guide<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/publisher-china-telecom-en\/\">China Telecom (Xing, TeleChat): models and licenses<\/a><\/td>\n<\/tr>\n<tr>\n<td>Published<\/td>\n<td>2026-10-16<\/td>\n<\/tr>\n<tr>\n<td>License<\/td>\n<td>apache-2.0<\/td>\n<\/tr>\n<tr>\n<td>Formats<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/format-safetensors-en\/\">safetensors<\/a><\/td>\n<\/tr>\n<tr>\n<td>Paper<\/td>\n<td><a href=\"https:\/\/arxiv.org\/abs\/2512.24157\">arXiv:2512.24157<\/a>, <a href=\"https:\/\/arxiv.org\/abs\/2507.18013\">arXiv:2507.18013<\/a><\/td>\n<\/tr>\n<tr>\n<td>Source type<\/td>\n<td>Primary source (the publisher itself)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><em>Values determined by this site&#8217;s code when the information was collected. Dates are JST.<\/em><\/p>\n<p><!-- \/lmw:facts --><\/p>\n<h2>Overview<\/h2>\n<p>&#8220;Xing4.0-29B-A4B-w8a8&#8221;, the W8A8 quantized version of &#8220;Xing4.0-29B-A4B&#8221; developed by China Telecom Artificial Intelligence Technology Co., Ltd., has been released. By quantizing both weights and activations to 8-bit integers, the model reduces memory usage while maintaining accuracy and improves inference efficiency on Ascend NPU devices. This model is a MoE model that activates only 4B out of a total of 29B parameters per token, natively supporting a 256K context length with a maximum of 512K.<\/p>\n<p>The paper associated with this model (<a href=\"https:\/\/arxiv.org\/abs\/2512.24157\">arXiv:2512.24157<\/a>) presents a training infrastructure that enables stable and efficient scaling to large model sizes on Ascend NPU clusters. Through a systematic parallelization framework including operator-level and end-to-end numerical precision verification, pipeline scheduling, and communication optimization, it aims to establish a foundation for large language model development within the hardware ecosystem.<\/p>\n<h2>Specifications<\/h2>\n<ul>\n<li>Total Parameters: 29B<\/li>\n<li>Active Parameters: 4B per token<\/li>\n<li>Architecture: mHC + MLA + MTP architecture (MoE)<\/li>\n<li>Layers: 40<\/li>\n<li>Hidden Size: 3584<\/li>\n<li>Dense FFN Intermediate Size: 9216<\/li>\n<li>Expert Intermediate Size: 1024<\/li>\n<li>Attention Type: MLA<\/li>\n<li>Routing Experts: 64<\/li>\n<li>Active Experts per Token: 4<\/li>\n<li>Shared Experts: 1<\/li>\n<li>Context Length: 256K (extensible up to 512K)<\/li>\n<\/ul>\n<h2>Performance<\/h2>\n<p>A comparison table of benchmark scores published in the model card is shown below. Open-weight models of comparable scale, Gemma4-26B-A4B and Qwen3.6-35B-A3B, are listed as comparison targets. Note that these figures are measurement results published by the authors themselves.<\/p>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Benchmark<\/th>\n<th style=\"text-align: center;\">Xing4.0-29B-A4B<\/th>\n<th style=\"text-align: center;\">Gemma4-26B-A4B<\/th>\n<th style=\"text-align: center;\">Qwen3.6-35B-A3B<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>IFBench<\/td>\n<td style=\"text-align: center;\">69.67<\/td>\n<td style=\"text-align: center;\">72.67<\/td>\n<td style=\"text-align: center;\">65.50<\/td>\n<\/tr>\n<tr>\n<td>AIME2026<\/td>\n<td style=\"text-align: center;\">90.00<\/td>\n<td style=\"text-align: center;\">88.30<\/td>\n<td style=\"text-align: center;\">92.70<\/td>\n<\/tr>\n<tr>\n<td>AA.LCR<\/td>\n<td style=\"text-align: center;\">61.00<\/td>\n<td style=\"text-align: center;\">66.00<\/td>\n<td style=\"text-align: center;\">62.00<\/td>\n<\/tr>\n<tr>\n<td>Tau3-Bench<\/td>\n<td style=\"text-align: center;\">64.63<\/td>\n<td style=\"text-align: center;\">58.90<\/td>\n<td style=\"text-align: center;\">67.20<\/td>\n<\/tr>\n<tr>\n<td>Claw-Eval<\/td>\n<td style=\"text-align: center;\"><strong>76.55<\/strong><\/td>\n<td style=\"text-align: center;\">71.49<\/td>\n<td style=\"text-align: center;\">74.54<\/td>\n<\/tr>\n<tr>\n<td>SWE-bench Verified<\/td>\n<td style=\"text-align: center;\"><strong>75.00<\/strong><\/td>\n<td style=\"text-align: center;\">53.00<\/td>\n<td style=\"text-align: center;\">76.00<\/td>\n<\/tr>\n<tr>\n<td>Terminal-Bench 2.1<\/td>\n<td style=\"text-align: center;\"><strong>57.50<\/strong><\/td>\n<td style=\"text-align: center;\">30.00<\/td>\n<td style=\"text-align: center;\">51.50<\/td>\n<\/tr>\n<tr>\n<td>SWE-bench Multilingual<\/td>\n<td style=\"text-align: center;\">66.00<\/td>\n<td style=\"text-align: center;\">51.00<\/td>\n<td style=\"text-align: center;\">67.20<\/td>\n<\/tr>\n<tr>\n<td>DeepresearchBII<\/td>\n<td style=\"text-align: center;\"><strong>60.80<\/strong><\/td>\n<td style=\"text-align: center;\">39.30<\/td>\n<td style=\"text-align: center;\">59.70<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>These measurement results show that Xing4.0-29B-A4B exhibits high aptitude in agent capabilities, device operation, and research-oriented tasks. In particular, it recorded 57.50 on Terminal-Bench 2.1, which measures whether commands can be actually executed in a terminal to complete given tasks, significantly outperforming Gemma4-26B-A4B&#8217;s 30.00 and Qwen3.6-35B-A3B&#8217;s 51.50 to achieve the highest score. It also achieved the highest figures among the comparison targets in Claw-Eval (76.55) and DeepresearchBII (60.80).<\/p>\n<p>On SWE-bench Verified, which evaluates whether an issue in an actual GitHub repository can be resolved by editing the repository, it shows a high level at 75.00, though falling slightly short of Qwen3.6-35B-A3B&#8217;s 76.00. Furthermore, in categories such as Tau3-Bench (64.63), which measures the ability to balance dialogue and tool invocation while adhering to business procedures, IFBench (69.67), and AA.LCR (61.00), it trails behind Gemma4-26B-A4B and Qwen3.6-35B-A3B, resulting in strengths and weaknesses depending on the nature of the task.<\/p>\n<h2>Strengths and Use Cases<\/h2>\n<p>Xing4.0-29B-A4B-w8a8 (base model: Xing4.0-29B-A4B) is designed with an emphasis on agent capabilities and the automated execution of complex tasks. Built with the mHC + MLA + MTP architecture, it excels at multi-step planning, external tool invocation, and the execution of complex reasoning chains. Even under ultra-long-context environments of 256K (extensible up to 512K), it maintains task consistency and execution stability.<\/p>\n<p>Compatibility with external tools and frameworks has also been considered, with format adjustments and target adaptations implemented for various agent frameworks such as OpenCode, Claude Code, OpenClaw, and Hermes. This enables smooth integration into existing agent systems and development workflows.<\/p>\n<p>Additionally, it is crafted to facilitate domain-specific tuning using specific domain data. Through lightweight fine-tuning with proprietary data, specialized vertical domain models can be built in a short time and at low cost for tasks such as the following:<\/p>\n<ul>\n<li>Intent Classification<\/li>\n<li>Table Understanding<\/li>\n<li>Contract Auditing<\/li>\n<li>Knowledge-based QA<\/li>\n<\/ul>\n<p>Supported fine-tuning frameworks include LLaMA-Factory and MindFormers.<\/p>\n<p>Regarding inference settings, recommended generation parameters according to task characteristics are provided:<\/p>\n<ul>\n<li>Complex reasoning \/ General tasks: temperature 1.0, top_p 0.95, repetition_penalty 1.05<\/li>\n<li>Coding \/ Agent tasks: temperature 0.8, top_p 0.95, repetition_penalty 1.05<\/li>\n<\/ul>\n<p>In addition, enabling and disabling the thinking process can be controlled by toggling the <code>enable_thinking<\/code> option within <code>chat_template_kwargs<\/code> when making calls via the OpenAI-compatible API.<\/p>\n<h2>How to Get It<\/h2>\n<p>The model &#8220;Xing4.0-29B-A4B-w8a8&#8221; is available in safetensors format on the Hugging Face repository. It is released under the Apache 2.0 license and can be downloaded without requiring license agreement procedures (gated).<\/p>\n<p>Quantizing both weights and activations to 8-bit integers (W8A8), this model is optimized for inference execution on Ascend NPU devices. The following tools are supported as engines for model deployment and serving:<\/p>\n<ul>\n<li>vLLM<\/li>\n<li>SGLang<\/li>\n<li>KTransformers<\/li>\n<\/ul>\n<p>It also supports connection via OpenAI-compatible API endpoints and can be called and used using Python&#8217;s <code>openai<\/code> library as follows:<\/p>\n<pre><code class=\"language-python\">from openai import OpenAI\n\nclient = OpenAI(\n    base_url=&quot;your-base-url&quot;,\n    api_key=&quot;your-api-key&quot;,\n)\n\ncompletion = client.chat.completions.create(\n    model=&quot;Xing4.0-29B-A4B&quot;,\n    messages=[{&quot;role&quot;: &quot;user&quot;, &quot;content&quot;: &quot;Briefly explain the basic principles of quantum computing.&quot;}],\n    temperature=1.0,\n    top_p=0.95,\n    extra_body={\n        &quot;repetition_penalty&quot;: 1.05,\n        &quot;skip_special_tokens&quot;: False,\n        &quot;spaces_between_special_tokens&quot;: False,\n        &quot;chat_template_kwargs&quot;: {\n            &quot;enable_thinking&quot;: True,\n        },\n    },\n)\n\nprint(completion.choices[0].message.content)\n<\/code><\/pre>\n<p>Specific serving procedures and detailed startup code are summarized in the official GitHub repository (<a href=\"https:\/\/github.com\/XingChen-AGI\/Xing4.0-29B-A4B\">https:\/\/github.com\/XingChen-AGI\/Xing4.0-29B-A4B<\/a> ).<\/p>\n<p><!-- lmw:runnability --><\/p>\n<h2>Can You Run It Locally?<\/h2>\n<p><strong>Not usable in Ollama, LM Studio and llama.cpp yet \u2014 we have found no GGUF build.<\/strong><\/p>\n<p>The publisher ships safetensors only. However, llama.cpp&#8217;s registry does list this architecture, so <strong>conversion to GGUF is possible<\/strong> and the model will run once someone publishes a converted build. Today it can be run with transformers, using the memory figures in the table above.<\/p>\n<p><strong>License \u2014 <code>apache-2.0<\/code> (Commercial use allowed):<\/strong> Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.<\/p>\n<p><em>Compiled by this site&#8217;s code from the published formats, converted builds we have found, and each engine&#8217;s own model registry. &#8220;Not found&#8221; means we have not seen such a build, not that none exists. License summaries are not legal advice \u2014 check the publisher&#8217;s original terms before relying on them.<\/em><\/p>\n<p><!-- \/lmw:runnability --><\/p>\n<p><!-- lmw:related --><\/p>\n<h2>Related Articles<\/h2>\n<ul>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/09\/28\/xing4-0-29b-a4b-china-telecom\/\">Xing4.0-29B-A4B Text Generation Model: Our Test Answers, 24GB+ VRAM<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/kimi-k3-1-pro\/\">Kimi-K3.1-Pro 1M-Context Multimodal Agent Model: ~1128GB Memory<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/ling-3-0-flash-sante-fp8-released\/\">Ling-3.0-flash-Sante-fp8 Text Generation Model: ~144GB Memory<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/ling-3-1-flash-1019-released\/\">Ling-3.1-flash-1019 Text Generation Model: ~641GB Memory, GGUF Builds<\/a><\/li>\n<\/ul>\n<p><!-- \/lmw:related --><\/p>\n<p><!-- lmw:next-steps --><\/p>\n<h2>What to Read Next<\/h2>\n<ul>\n<li><strong>Formats this model is available in<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/format-safetensors-en\/\">Safetensors format guide and models<\/a><\/li>\n<li><strong>Learn about the publisher<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/publisher-china-telecom-en\/\">China Telecom (Xing, TeleChat): models, licenses and articles<\/a><\/li>\n<li><strong>Other models for the same task<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/models-by-task-en\/#task-text\">Other text generation models<\/a><\/li>\n<li><strong>How to read SWE-bench, Terminal-Bench<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/glossary-benchmarks-en\/\">Benchmark glossary<\/a><\/li>\n<\/ul>\n<p><!-- \/lmw:next-steps --><\/p>\n<h2>Sources<\/h2>\n<ul>\n<li><a href=\"https:\/\/huggingface.co\/XingChen-AGI\/Xing4.0-29B-A4B-w8a8\">XingChen-AGI\/Xing4.0-29B-A4B-w8a8 (Hugging Face)<\/a><\/li>\n<li><a href=\"https:\/\/arxiv.org\/abs\/2512.24157\">Training Report of TeleChat3-MoE (arXiv)<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>Xing4.0-29B-A4B-w8a8, an 8-bit quantized MoE model by China Telecom, is now available. Features 29B total parameters, 256K context, and NPU optimization.<\/p>\n","protected":false},"author":1,"featured_media":12080,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[310],"tags":[165,169,1547,592,3614,2553,3616,2555,1555],"class_list":["post-12081","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-new-models","tag-moe-en","tag-sglang-en","tag-verified","tag-vllm-en","tag-w8a8-en","tag-xing4-0-29b-a4b-en","tag-xing4-0-w8a8-en","tag-xingchen-agi-en","tag--en"],"lang":"en","translations":{"en":12081,"ja":12079},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/12081","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/comments?post=12081"}],"version-history":[{"count":2,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/12081\/revisions"}],"predecessor-version":[{"id":17835,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/12081\/revisions\/17835"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media\/12080"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=12081"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/categories?post=12081"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/tags?post=12081"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}