Xiaomi Open-Sources MiMo-V2.6 as Alibaba Expands Qwen Agent Platform and Jev Offers 120 Million Tokens

MiMo-V2.6 comes in two versions. MiMo-V2.6-Pro has 1.02 trillion total parameters and 42 billion active parameters; MiMo-V2.6-Flash has 309 billion total parameters and 15 billion active parameters. Leiphone reported that Xiaomi devoted more attention to post-training than to raw scale. The company placed code, general Agent, vision and cybersecurity tasks into a single reinforcement-learning system. A single RL update uses 1,568 prompts and generates 16 rollouts per prompt, or 25,088 trajectories. Unlike ordinary question answering, Agent tasks require the model to read the environment, search for information, call tools, execute actions and adjust the next step based on returned results. The training target is therefore not only the final answer but an entire behavior trajectory containing states, decisions and feedback.

To support long-horizon generation, MiMo-V2.6-Pro uses a 70-layer Transformer, with 60 layers using sliding-window attention with a 128-token window and 10 layers using global attention. Each MoE layer has 384 routed experts, of which 8 are activated per token. Xiaomi also added a five-layer MTP speculative decoder that predicts the next seven tokens in one forward pass. Leiphone reported that the combination is intended to keep large capacity while controlling the per-step cost of sparse computation, local attention and speculative decoding. The training system uses fully asynchronous GRPO, separating rollout generation, environment execution, grading and model updates. Xiaomi mixed multiple domains and Agent harnesses into one RL process, which it calls You Only RL Once.

The report said Xiaomi redesigned reward mechanisms because binary pass-or-fail rewards become too coarse for long Agent trajectories. Groupwise Reward Synthesis, or GRS, observes multiple rollouts from the same task, builds a task-specific rubric and combines process quality with test results. Groupwise Advantage Redistribution, or GAR, continues to compare trajectories even when several complete the task, assigning more advantage to higher-quality solutions. Xiaomi said the process uses environment hardening, anomaly screening and verifier cross-checking to address reward hacking. After mixed RL, Xiaomi added MOPD2, or Multi-Prefix Multi-Teacher On-Policy Distillation, which reuses Teacher trajectories and SFT demonstrations by cutting a trajectory at an intermediate state and letting the student continue from there. The report said this can turn one teacher trajectory into multiple reusable training starting points.

At the MaaS and Agent main forum of the 2026 Yunqi Conference on September 22, Alibaba Group strategic vice president and ATH business group MaaS business line president Wen Zheng announced a full upgrade of the Qwen AI platform, QbitAI reported. The platform added Agent services and industry AI solutions on top of model services, and launched an API fast mode, Agent Studio and a Qwen AI cabin solution. Wen said AI remains early and that the key question is who can convert intelligence into results. Alibaba Cloud's MaaS platform customer count grew sixfold in the past year. Alibaba said it will keep investing in full-stack AI and expects the computing power operated by Alibaba Cloud to exceed 20 gigawatts by 2032.

On model services, the Qwen AI platform uses FlashBoot and UniScheduler to schedule heterogeneous computing power. QbitAI reported that elastic model startup time fell from 1,200 seconds to 70 seconds, 10,000 pods can be launched within one minute, first-token latency dropped 38 percent and Prompt Caching costs can be cut by up to 95 percent. The platform offers a 99.9 percent production SLA, billion-level peak TPM for a single customer and CMaaS confidential inference. The new API fast mode raises TPS by 1.5 to 2 times; throughput reservation adds an eight-hour night specification; dedicated throughput is aimed at large enterprises and financial or government customers. Token Plan covers Qwen, Kimi, GLM and DeepSeek models and connects to Qoder, Cursor, Codex and OpenClaw, with shared deductions through Qwen App and Qwen Office.

Agent Studio is an enterprise full-stack Agent service platform that automatically routes tasks to models, provides 24-hour managed operation and supports internal knowledge data and external software services. It offers more than 50 atomic APIs covering runtime, long-term memory, tool services, session requests and service endpoint deployment. An Agent API is coming, and an Agent self-evolution engine will include runtime observation, AI evaluators, AI optimizers and automatic verification. One Key MCP uses one API key to connect more than 100 ecosystem services, including Amap, Fliggy, 1688 and services in finance and legal fields. QbitAI reported that Meoo 1.0, Lingque 2.0, Wanjing Yike and Wan Xiaozhi 3.0 are among the platform's own products and applications. Lingque 2.0 serves 20 industries, more than 5,000 enterprise customers and over 100,000 seats. Industry users include Fanhai Group, NetEase Games, Honor and BYD, the report said.

Leiphone also reported that Jev, after drawing attention for speed and low cost, granted approved users 120 million tokens. The report questioned the move because Jev is already highly token-efficient. It mainly handles classification, routing, scoring and probabilistic judgment, and a short text input may use only dozens of tokens. At that rate, 120 million tokens could support millions of judgments. The report said Jev is not selling generated content but a single judgment. In Agent workflows, Jev is positioned at the System 1 layer: frequent small decisions such as intent recognition, tool selection, result filtering, exception handling and state checks. Calibration, or probability calibration, determines whether a judgment is confident enough to execute or should be passed to a stronger model. The report said large-scale real use can expose edge cases that lab experiments miss and help improve calibration, though it added that Jev's actual data policy would determine whether user requests enter training. The report connected the giveaway to the Jevons paradox: as the unit cost of judgment falls, developers may create new call points and total consumption may rise. Jev's aim, according to Leiphone, is to become the default quick-decision layer in Agents.