4月国产大模型格局重塑:GLM 5.1 领跑、Kimi K3 官宣、DeepSeek V4 压轴

4月国产大模型格局重塑:GLM 5.1 领跑、Kimi K3 官宣、DeepSeek V4 压轴

Core Assessment

April 2026 has emerged as a “month of collective breakthrough” for Chinese LLM developers. Starting with the impressive programming capabilities of GLM 5.1 at the beginning of the month, followed by Xiaomi’s open-sourcing of the mimo v2.5 series in mid-month, and culminating in the release of DeepSeek V4’s trillion-parameter MoE architecture at the end of April, the capability gap among domestic models has narrowed significantly within a single month. The latest LM Arena data shows Wenxin 5.1 Preview firmly holding the top spot among domestic models and ranking 13th globally, with DeepSeek V4 Pro at 23rd and Xiaomi mimo-v2.5-pro at 22nd. For the first time, multiple domestic models have simultaneously entered the global top 25 in the same period.

Key Release Timeline for April

DateModel/EventKey Features
Early AprilZhipu GLM 5.1Programming capabilities enter the “entry” tier; hands-on evaluation ≈ Kimi K2.6
Mid-AprilMoonshot Kimi K2.6 Open-SourcedOpen-sourcing a coding model caused industry ripples; standout multi-model Agent capabilities
Late AprilXiaomi mimo v2.5 SeriesOpen-source + multi-language/dialect ASR + token efficiency optimization
April 30DeepSeek V4 Pro Limited 75% DiscountTrillion-parameter MoE model API price hits a new low
May 1Moonshot Officially Announces Kimi K32.5 trillion parameters, Q3 release, directly targeting top-tier international models
OngoingWenxin 5.1 Preview#1 among domestic models on LM Arena, 13th globally

Hands-On Evaluation Matrix

Cross-validated by multiple independent developers. While not an official benchmark, it reflects the performance ranking in real-world usage scenarios:

TierModelTypical Scenario Performance
Above EntryGLM 5.1 ≈ Kimi K2.6Complex coding tasks, long-context reasoning
Above EntryDeepSeek V4 ProMost cost-effective large-parameter model
Below EntryQwen 3.6 Max PreviewWell-rounded overall, but slightly weaker in coding
Below Entrymimo v2.5 Pro > Qwen 3.6 PlusStrong performance in specific scenarios

Interpreting the Landscape Shift

A “4+1” structure is forming in the top tier: GLM 5.1, Kimi K2.6, DeepSeek V4 Pro, and Wenxin 5.1 Preview have reached the same tier in programming and comprehensive capabilities. With Xiaomi’s mimo v2.5 Pro as a fast follower, the “generational gap” among domestic models is rapidly disappearing.

Clear divergence in open-source strategies: Kimi K2.6 and Xiaomi mimo v2.5 have chosen the open-source route, while GLM 5.1 and Wenxin 5.1 continue to focus primarily on API/cloud services. DeepSeek, meanwhile, is adopting a “limited-time discount” strategy for V4 Pro to attract developer adoption.

The ambition behind K3: Riding the momentum of K2.6’s open-source release, Moonshot directly announced K3. Its 2.5 trillion-parameter scale signals a shift away from the “small but efficient” differentiation strategy, opting instead to compete head-on in the parameter race against top-tier international models.

Actionable Recommendations

  • Top choice for coding: GLM 5.1 or Kimi K2.6; both rank in the “entry” tier based on hands-on evaluations.
  • Best value option: The limited 75% discount period for DeepSeek V4 Pro (until May 5 / May 31) offers a low-cost window to experience a trillion-parameter MoE architecture.
  • Keep an eye on K3: Following its Q3 release, the domestic model landscape could be reshuffled again. Maintaining a wait-and-see approach is advisable.
  • Open-source ecosystem: Xiaomi mimo v2.5 series’ open-source strategy is worth monitoring, particularly its token efficiency optimizations, which offer valuable insights for edge deployment.