Skip to content
ES

Less noise. More signal.

Archived story — belongs to the Monday, November 2, 2026 sweep.
P1 · Tectonic · Frontier AI · sweep Monday, November 2, 2026

Open-weight release of Chinese models and inference cost compression

A bloc of Chinese laboratories (Xiaomi, MiniMax, Ant Group, ModelBest, NaiveAI) releases open weights of frontier-scale MoE models with sparse attention architectures that reduce the cost of long context, while Xiaomi trains a 1T-parameter model for ~$3M. The dynamic shifts competitive advantage from proprietary training toward architectural efficiency and open distribution, eroding the capital moat of Western laboratories and moving competition to the compute and energy layer required to serve these models.

Title, thesis and gravity assessment written by an AI model, without prior human review.

Gravity82/100 · Tectonic
Why this gravity
Xiaomi releases MiMo-V2.6 Pro/Flash weights with performance reported as similar to Opus 5 and GPT-5.6 Sol on agent benchmarks, and details HySparse2 that reduces prefill FLOPs for 1M tokens ~5x; the block of Chinese labs (MiniMax, Ant Group) replicates the opening of frontier MoE with sparse attention, eroding the capital moat of Western labs.
— Assessment by the Semantic Gravity Engine
Coverage
4 feeds · 4 independent publisher groups (corroborated)
Cross verticals
Frontier AI, Tech Industry, Energy & Materials, Geopolitics
Gravity vs. the previous edition
+15 (67 → 82)

coverage, not accuracy

First detected Thu, Aug 27 · 68 days trackedflagged as emergentGravity evolution
Lead source
Xiaomi open-sources MiMo-V2.6 models after scaling reinforcement learning · TechNode / TechNode Feed
Actors: Xiaomi · MiniMax · Ant Group · ModelBest · NaiveAI · SoMark
Timeline · Madrid time
  1. 09:52
  2. 00:10
  3. 04:01
  4. 08:30
  5. 09:06
  6. 04:00
  7. 07:57
  8. 06:04
  9. 06:24
  10. 06:47
  11. 08:49
  12. 09:17
  13. 15:51
  14. 02:52
  15. 03:54
ShareWhatsApp