AAITopic

← Back to today's topics

Verified · Jun 30, 2026

Chinese flagships line up on the same day: MiniMax M3 and Z.AI GLM-5.2

5 sources

MiniMax tags M3 as 'NEW' in its homepage nav, describing it as powered by MiniMax Sparse Attention (MSA), supporting 1M tokens of context (minimum 512K), natively multimodal, with a self-reported BrowseComp of 83.5 vs Opus 4.7's 79.3. Z.AI positions GLM-5.2 with a self-reported 51 on the Artificial Analysis composite (top three alongside Anthropic and OpenAI, open-source SOTA) and pushes GLM-5V-Turbo as a multimodal Coding model. Moonshot's Kimi homepage marks K2.6 as the flagship alongside Claw / Code / Doodle. Qwen frames Qwen Studio as a unified product surface.

Why now

On the same capture day (2026-06-30), four Chinese vendors all lined up their flagship model and product-surface positioning — a rare round-up moment creators can write as one piece instead of four loose posts.

Why it is worth publishing

This is one of the rare captures where MiniMax, Z.AI, Kimi, and Qwen all expose their current flagship positioning in one sweep. Creators can split 'flagship model' vs 'product surface' and break out MSA / 1M context / SOTA / multimodal Coding / Agent cluster / Qwen Studio cleanly.

Evidence basis

'China is matching frontier labs' is a perennial creator topic. Four flagships surfacing in one day makes for a clean round-up.

“MiniMax just put M3 on its homepage nav, and Z.AI is pushing GLM-5.2 at the same time.”

Angle

Split the four vendors into 'flagship model' (MiniMax M3 / Z.AI GLM-5.2 / Kimi K2.6) and 'product surface' (Z.AI GLM-5V-Turbo / Qwen Studio / Kimi Claw-Code-Doodle), not a vendor-by-vendor bullet list.

Format

Carousel

Demo idea

Build a 4-row × 2-column card. Left column = flagship model (M3 / GLM-5.2 / K2.6 / Qwen Studio). Right column = product pitch (MSA + 1M / Artificial Analysis 51 + SOTA / Claw-Code-Doodle / unified multimodal surface). Tag every row 'vendor self-statement, homepage meta'.

Platform notes

Always label the publishedAt as 'homepage meta capture' — the actual launch date is not disclosed in any of the four vendor meta descriptions. BrowseComp 83.5 / Artificial Analysis 51 must be marked as vendor self-reported numbers, not independently verified. When quoting 'SOTA' / 'frontier' / 'lossless', label them as 'vendor positioning' rather than as independent benchmark results. Do not invent a specific Qwen3 / Qwen3.5 / Qwen3-Max version number for the Qwen Studio summary.

Usable claims

  • MiniMax's homepage navigation (2026-06-30 fetch) tags MiniMax M3 as 'NEW' alongside the existing M2.7 and M2.5, listing it as the current top-level text model.
  • MiniMax's M3 model page (2026-06-30 fetch) describes MiniMax M3 as powered by MiniMax Sparse Attention (MSA) supporting up to 1M tokens of context with a guaranteed minimum of 512K tokens, with native multimodal training from step zero.
  • MiniMax's M3 model page reports a BrowseComp score of 83.5 for M3, surpassing Opus 4.7 (79.3), framed as evidence of strong autonomous browsing and information retrieval.
  • Z.AI's homepage (2026-06-30 fetch) states that GLM-5.2 scored 51 on the Artificial Analysis composite benchmark, placing it in the top three alongside Anthropic and OpenAI and at SOTA among open models.
  • Z.AI's homepage describes GLM-5.2 as open-source SOTA on Coding with 1M lossless context and more stable long-horizon execution.
  • Z.AI's homepage frames GLM-5V-Turbo as a natively multimodal Coding model that fuses vision and text, optimized for visual programming and 龙虾 (OpenClaw-style) agent tasks.
  • Moonshot's Kimi homepage (2026-06-30 fetch) positions Kimi K2.6 as the current flagship, with adjacent products Kimi Claw / Kimi Code / Kimi Doodle and an upgraded Agent cluster plus Office-to-Skill conversion.
  • Qwen's homepage (2026-06-30 fetch) advertises Qwen Studio as a single product surface spanning chatbot, image and video understanding, image generation, document processing, web search integration, tool utilization, and artifacts.

Evidence pipeline

From the news

Breakdown

This breakdown threads MiniMax M3 (MSA + 1M context + BrowseComp 83.5), Z.AI GLM-5.2 (Artificial Analysis 51 + 1M lossless context + Coding SOTA), GLM-5V-Turbo (multimodal Coding + OpenClaw / 龙虾 agent positioning), Moonshot Kimi K2.6 (flagship + Claw / Code / Doodle) and Qwen Studio (unified product surface) into two lines — flagship model and product surface — and flags that every number and quote is vendor-self-reported from the homepage meta, with the actual launch date not disclosed.

Sources

Risks

  • Mark each CN model signal's publishedAt as the AITopic capture date and add a risk note clarifying the actual launch date is not on the homepage.
  • Attribute '51 on Artificial Analysis' and 'top three' to Z.AI's homepage; pin to Artificial Analysis if a direct link exists.
  • Quote the 83.5 number as MiniMax's self-reported score; do not claim independent replication.
  • Quote capability phrases from official pages but mark them as 'vendor positioning', not as independent benchmarks.
  • Frame the Qwen signal as 'the current Qwen Studio surface' rather than a specific model version.

Demo ideas

  • Make 4 cards: architectural novelty (M3 MSA) / composite benchmark (GLM-5.2) / product cluster (K2.6 + Claw/Code/Doodle) / unified surface (Qwen Studio)
  • Compare GLM-5.2 '1M lossless' with MiniMax M3 '1M context (minimum 512K)' and explain the gap
  • Stack the four 'frontier' pitches as a single timeline tagged with each vendor's homepage meta capture day