Xiaomi Mimo 2.6 live post-training dashboard

Original Article ↗ Hacker News Discussion ↗

Xiaomi has launched a live dashboard showing post-training reinforcement learning runs for its MiMo 2.6 coding models, revealing real-time costs, progress, and benchmark scores in an industry that is typically opaque. Commenters praise the transparency and strong cost–performance of earlier MiMo models, compare them with competitors like DeepSeek, GLM, Qwen, and Gemini, and debate issues such as benchmark contamination, data composition, and possible distillation from frontier models. Some are skeptical about whether the dashboard reflects live data, but many see it as a notable contrast to the secrecy of large US labs and a sign of growing openness from Chinese AI companies.

Overall reaction to the dashboard

  • Many find the live post-training / RL dashboard “remarkably transparent” and “cool,” especially in a generally secretive industry.
  • Several users say they’re learning a lot about training dynamics just by watching the run.
  • Others see it as mainly a clever PR move rather than a fundamental shift.

Transparency, openness, and motives

  • Some wonder why other labs (OpenAI, Anthropic, Google, IBM) don’t do similar dashboards; proposed reasons:
    • Big US labs believe they have “special sauce” and gain little by exposing methodology.
    • Smaller/Chinese labs benefit more from openness as a differentiator and marketing.
  • It’s suggested that Chinese policy now explicitly favors open models and open development, creating incentives for visible openness.

Training data, RL, and benchmarks

  • Clarification that this is a post-training RL run, not pretraining, though some note even pretraining often uses 30–50% code.
  • Debate on whether running benchmarks during training is “contamination”:
    • One view: benchmarks act as validation/early-stopping signals, causing mild overfitting but far less than training directly on them.
    • Another view: even as stopping criteria, they implicitly tune toward those benchmarks and contribute to “benchmaxxing.”
  • Benchmarks are also framed as necessary to detect degradation or bad data batches.

Authenticity and reliability of the dashboard

  • A subset of commenters suspect the dashboard is partially or wholly fake (e.g., progress jumping backward on refresh, restarts not matching graphs, “ticker” behavior).
  • Others counter that intermediate tickers can be synthetic while periodic real data updates are still genuine, likening it to a progress bar.
  • Overall, whether the feed is fully real-time and faithful is viewed as unclear.

Model quality and benchmarks

  • DeepSWE scores for MiMo 2.6 look very high compared to MiMo 2.5 Pro’s reported 19%, putting it near “frontier” coding models in that benchmark.
  • Some argue DeepSWE is now saturated and less discriminative at the top end; still useful to distinguish mid vs top-tier models.

Pricing, ROI, and practical use

  • Multiple users report MiMo 2.5 (and 2.5 Pro) as strong coding workhorses with excellent cost-per-token, especially relative to US and other Chinese models.
  • Common pattern: use a more capable or expensive model for planning/design, then MiMo for obedient implementation or background tasks.
  • Comparisons are made against DeepSeek, GLM, Qwen, Gemini, etc., with tradeoffs of quality vs price vs speed.
  • Some see open-weight and cheaper models as increasingly attractive for enterprise: predictable performance, tunability, privacy, flat pricing with owned hardware.

Geopolitics and industry contrast

  • Several comments frame Chinese labs as “showing up” US labs by being more open and experimental, while US labs focus on safety rhetoric, closed models, and valuation.
  • Others push back, noting that many Chinese models are viewed as distillations of US “frontier” models, leading to debate over who truly leads and who copies.