Xiaomi has published a live post-training dashboard displaying real-time telemetry, compute expenditures, and evaluation metrics for its Mimo 2.6 reinforcement learning (RL) model pipeline. The dashboard indicates an RL training burn rate of approximately $5 per second (~$432,000 per day), with two-thirds of the post-training dataset consisting of source code. Early evaluation logs show Mimo 2.6 Pro reaching a 63.7% score on the DeepSWE 1.1 benchmark by step 10, representing a significant performance jump over Mimo 2.5 Pro's 19% baseline. The public interface offers rare visibility into enterprise post-training dynamics, triggering community discussion regarding lab transparency, benchmark evaluation methodology, and model cost-performance ratios.
Key Points
Live Post-Training Telemetry: Xiaomi exposed real-time reinforcement learning dashboard metrics for its upcoming Mimo 2.6 model series, tracking step progression, compute costs, and live evaluations.
Compute Expenditure and Scale: The Mimo 2.6 Pro RL training run began on 2026-09-15 at 10:32 UTC, accumulating over $1.2 million in total cost at an operating rate of $5 per second ($432,000/day), equivalent to renting roughly 3,000 B300 GPU nodes on Modal.
Code-Centric Data Mix: Source code comprises approximately 66% (two-thirds) of the overall post-training reinforcement learning dataset.
DeepSWE 1.1 Performance Leap: Mimo 2.6 Pro achieved a 63.7% score on DeepSWE 1.1 by step 10, while Mimo 2.6 Flash reached 60.7% at step 12, heavily outperforming Mimo 2.5 Pro's previous baseline score of 19%.
Frontier Baseline Context: Top-performing frontier models on the DeepSWE 1.1 benchmark at maximum effort include Astra (74%), Fable (70%), and Kimi K3 (69%).
Discussion Highlights
Cost-Efficiency & Real-World Utility: Engineers report high return on investment using Mimo 2.5 Pro as an automated coding and testing agent, noting that while it exhibits conservative code generation, its output quality rivals legacy Anthropic models at an order-of-magnitude lower token cost.
Developer Workflow & Next-Gen Capabilities: Early testers of Mimo 2.6 note substantial improvements in multi-tasking, proactive solution design, and user interface translation compared to Mimo 2.5 Pro's minimal and slightly forgetful implementation style.
Comparative Model Ecosystem: Community members benchmarked Mimo against alternatives, noting that while Qwen 3.8-Flash-Next, DS 4.1 Flash, and GLM 5.3 Flash excel in specific local or terminal tasks, Mimo remains highly competitive due to ultra-fast inference options like the 1,000 token/sec UltraSpeed mode.
Benchmark Contamination Debate: Participants debated whether running benchmark evaluations during RL loops introduces data contamination; consensus noted that using benchmarks for validation-set tracking and early stopping creates mild selection bias ("benchmaxxing") but does not equal direct training contamination.
Transparency vs. Authenticity: Commenters praised Xiaomi and open Chinese AI initiatives for public training transparency over secretive US labs (OpenAI, Anthropic), though some skeptics questioned dashboard authenticity, alleging metric anomalies across browser refreshes and process restarts.