From Spark-X2.5 to Spark-X2.6: Cloud-Teacher Training, Offline Models, and Browser-Managed Local AI

[lostlight530]

English

On September 20, I shared an idea about making Spark-X2.5 available as a browser-integrated or browser-managed on-device model in Discussion #12

After the release of Spark-X2.6, I think this direction has become even more interesting, especially with the introduction of cloud-edge collaborative training and inference

One detail that caught my attention at the 2026 iFLYTEK 1024 Developer Festival was the training approach behind Spark-X2.6-4B: the smaller on-device model can be improved through adaptive post-training that selects teacher models and learning content according to the student model's capabilities

This suggests an interesting division of responsibilities:

Cloud-side teacher models help improve smaller models during training

On-device models handle tasks locally, including offline operation where supported

Cloud services remain available for tasks exceeding local capabilities, subject to user choice and appropriate data controls

Browsers and agent runtimes could make these capabilities accessible without requiring ordinary users to manage inference frameworks themselves

I have also recently completed an independent diagnostic experiment with the official XHToken/Spark-X2.6-4B checkpoint using BF16 inference on a Google Colab NVIDIA L4 GPU

The experiment covered 12 Chinese and English legal and financial reasoning cases, generating 45,171 tokens over approximately 3,009 seconds. All 12 cases passed the experiment's predefined automated checks

This was a cloud-based model evaluation, not a local deployment benchmark or a full professional legal/financial validation. Nevertheless, it provided useful hands-on experience with the model's inference behavior

For me, the interesting next question is no longer just whether a compact model can run locally. Spark-X2.6 already has an increasingly accessible local inference ecosystem

It is how to make that capability practical for more people

For example, a browser or lightweight agent runtime could manage model downloads, local inference, privacy settings, offline tasks and optional cloud fallback through a consistent interface

Once a model and its runtime are installed, properly configured local inference can continue without network access. This may reduce recurring cloud inference costs and allow sensitive content to remain on the device, although the actual privacy guarantees still depend on the application and runtime implementation

I would be interested in exploring how the cloud-teacher approach, compact on-device models, and browser-level model management could fit together

The goal would not be to replace cloud models, but to give users a more accessible, controllable and cost-conscious way to use local AI

This is still an engineering and product direction worth exploring, rather than a claim that such an integrated system has already been implemented

Relevant links:

Previous Spark-X2.5 browser discussion

Official Spark-X2.6 repository

Spark-X2.6-4B model

[lostlight530]

中文版本

9 月 20 日，我曾在 Discussion #12 提出过一个想法：能否让 Spark-X2.5 以浏览器内置或浏览器统一管理的方式，为普通用户提供端侧模型能力

随着 Spark-X2.6 发布，以及讯飞在 2026 年 1024 开发者节上介绍端云一体训练和推理协同架构，我觉得这个方向又有了新的讨论空间

其中我比较关注的是 Spark-X2.6-4B 的端云一体训练方法

公开介绍显示，端侧小模型可以通过能力感知的自适应迭代后训练，根据自身能力动态选择教师模型、调整学习内容，重点提升长文本、工具调用和智能体任务能力

这让我想到一个值得继续探索的体系：

云端教师模型负责帮助端侧小模型提升能力

端侧模型负责在本地处理适合自身能力的任务

对于本地难以完成的复杂任务，可以由用户决定是否启用云端协同

浏览器或轻量 Agent Runtime 则可以负责模型管理、本地推理、权限控制和使用入口

这样，模型的训练与实际使用可以处在不同的计算环境中

小模型可以借助云端训练体系获得能力提升，但用户在使用时，不必每次都依赖远程模型服务

最近我也独立完成了 Spark-X2.6-4B 的一次云端实测

使用官方 BF16 权重、Google Colab NVIDIA L4 和 Transformers，对中文与英文的法律、金融推理问题进行了 12 次诊断执行，累计生成 45,171 Tokens，推理总耗时约 3,009 秒，12 项均通过预设自动判定规则

这不是本地离线部署测试，也不能直接代表法律与金融领域的完整专业正确率，但让我进一步接触到了 X2.6 的实际推理表现

结合之前 X2.5 的本地量化运行研究，我现在更关注如何把这些模型能力真正交给普通用户

例如：

浏览器可以统一管理本地模型与推理服务，让用户直接使用中文网页问答、文档处理、隐私内容分析、轻量 Agent 和离线任务能力

在模型和运行环境已经准备好的前提下，离线推理可以不依赖网络请求，减少云端 API 的持续调用成本，并让需要本地处理的数据尽可能留在设备上

对于确实需要云端能力的任务，再通过明确的用户授权、数据边界和端云路由完成协同

我认为，云端教师模型提升小模型能力，与本地离线推理、浏览器统一管理并不矛盾，反而可能形成一条很有价值的技术和产品路径

Spark-X2.6 已经有开源权重和多种本地推理框架支持，未来值得继续探索的，是如何降低普通用户实际使用这些能力的门槛

这仍然只是基于模型体验、公开技术信息和独立实验提出的方向讨论，不代表已经实现完整的浏览器端云协同产品

也欢迎社区朋友分享对端侧模型、浏览器 AI、Agent Runtime，以及云端教师模型训练方式的看法