Xiaomi MiMo-V2.6 Launches: Open Weights, Radical Training Transparency

A launch that got more attention for what it disclosed than for what it claimed to beat. Xiaomi's MiMo team shipped MiMo-V2.6 on September 22, 2026 — open-weight Flash and Pro models plus a smaller distilled variant — five days after publicly livestreaming the model's reinforcement-learning training run. The launch hit 558 points on Hacker News, with the top comment praising Xiaomi's transparency rather than its benchmark scores — a genuinely unusual reaction for a model launch thread.

TL;DR — what people are actually asking

What actually shipped

Three open-weight checkpoints, each published with full weights on Hugging Face:

The distilled variant is a supervised fine-tune of Alibaba's Qwen3.5-9B on MiMo-generated data — a smaller, dense model aimed at builders who need something that fits on more modest hardware rather than the full mixture-of-experts checkpoints. Both Flash and Pro follow the MoE (mixture-of-experts) pattern that's become standard for frontier-adjacent open releases in 2026: a large total parameter count with a much smaller share actively computing on any given token, keeping inference cost manageable despite the headline size.

The transparency, not the benchmarks, is what won the thread

The top comment on the Hacker News thread — 5 hours old, well ahead of anything discussing raw scores — praised the realtime dashboard Xiaomi shared during training as "an incredible learning and teaching tool," specifically calling out how comprehensive the accompanying technical report was in documenting methodology, comparing it favorably to the level of detail typically seen in Google or DeepSeek writeups. A separate reply went further on specifics: Xiaomi logged every training restart with a stated reason, disclosed dropping a cyber dataset after finding it degraded coding benchmarks, and published real-time training loss, per-checkpoint benchmark scores, and running cost estimates — including scores on benchmarks where MiMo didn't come out ahead.

That last point is the one worth sitting with. Most model launches publish a curated benchmark table optimized to make the release look as strong as possible. A commenter explicitly noted MiMo's benchmark disclosure includes "stuff they didn't do well on" — a genuinely rare choice in a competitive model-launch cycle, and the specific behavior that generated calls in the thread for "if you're releasing an open model going forward, please consider offering the community more of this transparency."

One more technical detail drew attention from a commenter who dug into the dashboard: fields like partial/avg_staleness and train_infer_diff/new_infer/kl, which track how far the model being updated has drifted from the model generating its own training rollouts — a real, technically substantive on-policy RL detail rather than marketing copy, described as something the poster hoped to turn into a full blog post given the level of interest it generated in replies.

Where the benchmark claims actually stand

Not everyone in the thread was convinced by the launch's framing, and the skepticism was specific rather than reflexive. One commenter posted their own comparison table across three benchmarks, arguing plainly: "I don't trust any of the benchmarks where Opus 5 surpasses Astra or Fable 5.1. Maybe Terminal Bench 4.0 and ExploitGym are reasonable" — then showed MiMo-V2.6-Pro trailing GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5 on both Terminal-Bench 4.0 (34.9% vs. Astra's 59.6%, a wide gap) and ExploitGym (17.8% vs. Astra's 42.4%), while landing much closer to the pack on DeepSWE v1.1 (71.9%, essentially tied with Opus 5's 74.0% and Astra's 74.0%).

That's a meaningfully different picture than "frontier-beating open model." MiMo-V2.6's actual claimed edge, according to multiple commenters citing Artificial Analysis's own tracking, is intelligence-per-dollar, not raw capability — Pro reportedly sits on AA's Pareto frontier for intelligence versus cost at a small fraction of the price of similarly-scored closed and open competitors, and a separate commenter noted the 1-trillion-parameter Pro model outperforms the 2.8-trillion-parameter Kimi K3 on 14 of 15 published benchmarks — a genuinely notable efficiency result even if it doesn't put MiMo ahead of the absolute frontier on raw capability.

One more genuine point of praise, unrelated to the benchmark dispute: MiMo's frontend and interactive-task generation quality drew specific compliments for producing diverse task demonstrations (a DAW interface, scientific-environment usage, real-world tool use) rather than the narrower "chatbot answering a coding question" demo format most launches default to.

The bigger context: China's transparency-as-strategy pattern

Several commenters connected MiMo's approach to a broader observation about the Chinese open-model ecosystem in 2026 — that comprehensive methodology disclosure, live training visibility, and honest benchmark reporting have become a recognizable competitive strategy distinct from closed-lab practice, not a one-off choice by this particular team. One reply specifically tied this to Luo Fuli's background — she previously worked at Alibaba and DeepSeek before leading Xiaomi's MiMo effort — as a plausible explanation for why MiMo's go-to-market style echoes DeepSeek's own reputation for unusually detailed technical writeups.

This isn't a universally accepted framing. A more skeptical thread elsewhere in the comments argued that "no Chinese lab has caught up yet," alleging some competitors "fake it by distilling and overfitting on benchmarks," while a counter-reply pushed back that the "everyone catches up" dynamic in open models is a real, repeatable pattern that closed-model labs haven't found an answer to. That broader closed-vs-open economics debate is worth reading in full context rather than summarizing as settled either way — see the American closed AI vs. China's open-weights strategy debate for the fuller argument on both sides.

Real-world usage questions the launch didn't fully answer

Beyond the benchmark debate, a chunk of the discussion focused on practical deployment questions that the launch itself left open. One commenter directly asked whether people are actually running MiMo models as daily drivers in company settings, and if so, through what hosting path — noting that the available options (OpenCode's built-in provider, or calling Xiaomi's own API directly) both carry unclear data-training guarantees, since neither publishes a firm zero-data-retention commitment in the way some Western inference providers do. A reply pointed to a European-hosted token plan with a stated zero-data-retention policy and Amsterdam-based infrastructure as one workaround, but this remains a genuinely unresolved gap for teams evaluating MiMo for anything beyond personal or side-project use — pricing and raw capability aside, where the model actually runs and what data guarantees come with that hosting matters just as much for a production decision.

A separate informal but revealing signal came from outside the benchmark tables entirely: one commenter who ran MiMo, Qwen, and several other open models against each other playing social deduction and drawing games — with the models trash-talking each other live — found MiMo-V2.6's predecessor to be the strongest performer among models priced under $0.15 per million input tokens on OpenRouter specifically at game-playing tasks, a genuinely different evaluation axis than any formal benchmark table captures. That's weak evidence on its own, but it's consistent with the broader theme in the thread: MiMo's strength claims cluster around cost-adjusted, task-diverse performance rather than a clean frontier-benchmark win.

What this means for what you build or pay

• If cost-per-completed-task matters more than absolute frontier capability for your workload, MiMo-V2.6-Flash or Pro are worth benchmarking against your own tasks — the intelligence-vs-cost positioning is the actual, defensible claim here, not a frontier-beating one.
• If you're evaluating training methodology rather than just output quality, the public RL dashboard and technical report are genuinely worth reading regardless of whether you use MiMo itself — the on-policy RL staleness metrics and the disclosed-failure approach are useful reference points for anyone running their own RL fine-tuning.
• Don't take "beats Kimi K3 on 14 of 15 benchmarks" as "beats the frontier." That's an efficiency claim (smaller model beating a larger one), not a claim that MiMo leads GPT-6 Astra, Fable 5.1, or Opus 5 outright — the thread's own skeptics showed real gaps on Terminal-Bench 4.0 and ExploitGym specifically.

Related on explainx.ai

Update — September 22, 2026: MiMo-V2.6 has shipped — see this post for the final specs, benchmarks, and community reception, following the training-in-progress coverage below.

• Xiaomi MiMo-V2.6: Livestreaming a Trillion-Parameter RL Training Run
• American Closed AI vs. China's Open-Weights Strategy Debate
• Startups Are Switching to Open-Weight Models to Save Money
• Top Chinese AI Companies and Startups Guide
• US vs. Chinese AI Startups Comparison
• AI Benchmarks Complete Guide
• Claude Fable 5.1 / Mythos 5.1: Launch, Benchmarks, and Pricing

This post reflects a Hacker News launch thread (558 points) discussing Xiaomi's MiMo-V2.6 release, as of September 22, 2026. Benchmark figures cited from community comments (Terminal-Bench 4.0, ExploitGym, DeepSWE v1.1 comparisons) are third-party, unofficial numbers shared in the discussion, not an independently reproduced explainx.ai benchmark run. Follow @explainx_ai for updates.