Xiaomi Streams Agent RL Training Live

Read original on Pandaily

#agent-training#training-costsmimo-v2.6-pro/flash-rl-dashboardxiaomimimo-v2.6mimo

Xiaomi is exposing the steps, rewards, tokens, and costs behind large-scale agent reinforcement learning.

30-Second TL;DR

What Changed

The stream covers MiMo-V2.6 Pro and MiMo-V2.6 Flash.

Why It Matters

Public training telemetry could improve reproducibility and help practitioners understand the economics of post-training agents. It may also encourage other labs to disclose more operational metrics.

What To Do Next

Monitor the MiMo RL dashboard and record reward-per-token and cost trends to compare against your own agent post-training runs.

Who should care:Researchers & Academics

Key Points

•The stream covers MiMo-V2.6 Pro and MiMo-V2.6 Flash.
•Viewers can inspect steps, token usage, rewards, and cost.
•The dashboard offers uncommon visibility into agent RL at scale.

Key numbers$1 million$1.13 million$20,000$5

Deep Insight

Background and context from public sources — not the original article. 7 sources cited.

Enhanced Key Takeaways

•The training pipeline operates at a massive rollout scale, ingesting roughly 2 billion tokens across 1,568 prompts per step with 16 asynchronous rollouts per prompt.
•Real-time financial tracking on the dashboard revealed cumulative spending surpassing $1 million to $1.13 million within 36-48 hours, with peak Pro runs burning approximately $20,000 per hour ($5 per second).
•Live mid-training evaluations demonstrated that MiMo-V2.6-Pro reached 62% to nearly 66% on the DeepSWE v1.1 benchmark (mini-swe-agent, avg@3), up significantly from MiMo-V2.5's ~19% baseline.
•The initiative is led by Xiaomi MiMo lead Luo Fuli (formerly of DeepSeek), who confirmed plans to progressively open-source the technical recipes, details, and infrastructure tools.
•The agent RL effort runs strictly parallel to, but separate from, Xiaomi's concurrent Robotics-U0 embodied AI project (an autoregressive 38B/4B world model).

Competitor Analysis

Observability & Transparency

Xiaomi MiMo-V2.6 (Live RL Stream): Fully public live dashboard streaming metrics, spend, and restart logs in real time
DeepSeek-R1: Post-training recipes and weights released post-hoc via technical reports
MiniMax-M1: Closed training metrics; post-release reports/weights

RL Infrastructure Pipeline

Xiaomi MiMo-V2.6 (Live RL Stream): Fully asynchronous, decoupled rollout/grader pipeline across 5 agent environments
DeepSeek-R1: Asynchronous multi-stage RL framework
MiniMax-M1: Distributed RL fine-tuning setup

DeepSWE v1.1 Benchmark

Xiaomi MiMo-V2.6 (Live RL Stream): 62% - ~66% (mid-training score)
DeepSeek-R1: Not directly evaluated on DeepSWE v1.1
MiniMax-M1: Not directly evaluated on DeepSWE v1.1

Observed Post-Training Spend

Xiaomi MiMo-V2.6 (Live RL Stream): Surpassed $1M - $1.13M within 48h (~$20k/hr peak Pro burn)
DeepSeek-R1: Reported low single-digit million budget (not streamed live)
MiniMax-M1: Proprietary run budget (not streamed live)

Feature / Metric	Xiaomi MiMo-V2.6 (Live RL Stream)	DeepSeek-R1	MiniMax-M1
Observability & Transparency	Fully public live dashboard streaming metrics, spend, and restart logs in real time	Post-training recipes and weights released post-hoc via technical reports	Closed training metrics; post-release reports/weights
RL Infrastructure Pipeline	Fully asynchronous, decoupled rollout/grader pipeline across 5 agent environments	Asynchronous multi-stage RL framework	Distributed RL fine-tuning setup
DeepSWE v1.1 Benchmark	62% - ~66% (mid-training score)	Not directly evaluated on DeepSWE v1.1	Not directly evaluated on DeepSWE v1.1
Observed Post-Training Spend	Surpassed $1M - $1.13M within 48h (~$20k/hr peak Pro burn)	Reported low single-digit million budget (not streamed live)	Proprietary run budget (not streamed live)

Technical Deep Dive

Decoupled Asynchronous Rollout Pipeline: Operates fully asynchronously, decoupling rollout generation, environment execution, reward scoring, and parameter updates from rigid synchronization barriers.
Scale & Throughput: Ingests ~2 billion tokens per training step across 1,568 prompts, generating 16 asynchronous rollouts per prompt.
Multi-Task Agent Environments: Concurrently mixes five disparate agent environments—code, general reasoning, vision, chat, and cybersecurity—within a single unified training job.
Novel Credit Assignment & Grader Compute: Incorporates agentic in-group credit assignment driven by deterministic unit test cases paired with rubric-based grader scoring.

Future ImplicationsAI analysis grounded in cited sources

Frontier AI labs will face heightened pressure to open-source post-training metrics and live run economics.

Xiaomi's live exposure of failure logs, rollout counts, and financial burn sets an unprecedented public transparency standard that challenges the opaque, closed-door post-training practices of frontier competitors.

Xiaomi will emerge as a top-tier open-weight reasoning and agentic model provider.

The demonstrated leap on the DeepSWE v1.1 benchmark from ~19% to over 62% signals that MiMo-V2.6's upcoming open-source recipe could rapidly disseminate frontier agentic coding capabilities to the broader developer community.

Timeline

2026-03

Xiaomi MiMo team begins quiet experimentation on agent RL post-MiMo-V2.5

2026-09

Xiaomi launches public streaming dashboard for MiMo-V2.6 Pro and Flash RL training

2026-03
Xiaomi MiMo team begins quiet experimentation on agent RL post-MiMo-V2.5
2026-09
Xiaomi launches public streaming dashboard for MiMo-V2.6 Pro and Flash RL training