AI research wiki
ai-industry · 2026-05-03 · Tier 1

Xiaomi MiMo-V2.5-Pro — Open-Weight Long-Horizon Coding at 40-60% Fewer Tokens

Xiaomi MiMo-V2.5-Pro — Open-Weight Long-Horizon Coding at 40-60% Fewer Tokens

Source: The Decoder Raw: raw/rss/2026-05-03-the-decoder-xiaomis-open-weight-mimo-v25-pro-takes-aim-at-claude-op.md URL: https://the-decoder.com/xiaomis-open-weight-mimo-v2-5-pro-takes-aim-at-claude-opus-with-hours-long-autonomous-coding/ Date: 2026-05-03 Tier: 1/2 — open-weight long-horizon agent + token efficiency

TL;DR

Xiaomi released MiMo-V2.5-Pro, an open-weight model that nearly matches Claude Opus 4.6 on coding benchmarks while burning 40–60% fewer tokens per task (per company-reported numbers). Positioned for hours-long autonomous coding tasks. The framing in the article: the Chinese open-weight competition (DeepSeek, Xiaomi, Moonshot, Qwen) is shifting from "raw benchmark scores" to "how cheaply and how long a model can run autonomously on a single task."

Key claims

  • Near-parity with Claude Opus 4.6 on coding benchmarks (specific numbers in the post; The Decoder's RSS extract is the headline only)
  • 40–60% fewer tokens per task than the comparison model — the operative metric is now per-task cost, not capability ceiling
  • Hours-long autonomous coding focus — long-horizon agent positioning
  • Open-weight release

Why this matters for cere-bro (Tier 1)

This is the biggest open-weight competitive-pressure data point of the past 72 hours, and it lands directly on three Tier-1 threads:

  1. Tokens-per-task is the new pricing axis, not benchmark score. SemiAnalysis (05-01) argued AI lab pricing power persists because the closed-best/open-best gap is not closing on capability. Xiaomi's framing inverts this: they concede the capability ceiling and compete on per-task cost. If the 40–60% number generalizes beyond their internal evaluations, the pricing pressure on closed labs comes through efficiency, not ceiling. This is the structural pressure that SemiAnalysis's thesis did not anticipate.
  2. Long-horizon coding agents intersect Step-level Optimization (05-02). A model designed for hours-long autonomy is a model whose tokens-per-step matters. Whether MiMo-V2.5-Pro's efficiency comes from architecture (MoE? hybrid SSM?), training (on-policy distillation à la TIP/CoPD?), or trajectory shaping (LenVM-style length value heads, 05-01) is the open empirical question. The The Decoder post does not reveal the mechanism.
  3. Open-weight onshoring + Chinese frontier labs. Pairs directly with Moonshot/StepFun onshoring (05-01) and the broader China vs US benchmark debate (US benchmark says China is 8 months behind; Xiaomi's release is a counter-data-point on cost-efficiency). The bilateral AI sovereignty thread continues.

Open questions

  • Mechanism of 40–60% token reduction. No public technical detail in the headline. Candidates: MoE sparsity, length-aware reward shaping (LenVM 05-01 is open-weight reference), distillation from a stronger teacher (CoPD 05-01 paradigm), better tokenizer for code, RL-tuned chain-of-thought truncation. First independent third-party replication or technical post will resolve this.
  • Real-world long-horizon evaluation. "Hours-long autonomous coding" deserves Claw-Eval-Live (05-01) or InteractWeb-Bench (05-01) numbers. As of 2026-05-04 none published.
  • Comparison ceiling on multi-step composability. Lower per-step cost but higher per-step variance could blow up cumulative variance over hours. The trajectory-routing literature (Step-level Optimization 05-02) is the right framework to evaluate this.

Connections to prior wiki pages

  • SemiAnalysis AI Value Capture (05-01) — counterpoint. Xiaomi's release is the open-weight efficiency competition SemiAnalysis did not weight heavily.
  • LenVM token-level length value model (05-01) — candidate mechanism for token reduction.
  • CoPD Co-Evolving Policy Distillation (05-01) — candidate mechanism for capability transfer.
  • Step-level Optimization (05-02) — orthogonal axis. Xiaomi's reduction is per-step; Step-level Optimization is per-trajectory. Composing them is the deeper efficiency play.
  • China AI startups onshoring (05-01) — same week, same theme.
  • The Decoder: China is falling behind (05-03) — direct contradiction of the US-government-benchmark framing.

Worth Watching

  • Independent third-party benchmark in 30 days. Aider, SWE-Bench Verified, and Claw-Eval-Live (05-01) for token-cost-per-resolved-task. If 40–60% replicates, this is a new pricing baseline; if it doesn't, the figure was task-specific.
  • Mechanism disclosure in 60 days. A model card or technical report. The mechanism determines whether the efficiency is generalizable.
  • Closed-lab response in 90 days. If Anthropic or OpenAI ships an explicit "tokens-per-task" pricing tier or efficiency variant, the pricing axis has shifted. Tracker: any new Anthropic or OpenAI release framed around per-task cost rather than capability.