Topics

China AI model release tracker (how to keep the watchlist current)

Evergreen topic pages updated with new evidence

2026 Updated 6-lab coverage License-verified

Last reviewed: 2026-06-23 · Policy: Editorial standards · Methodology

Decision in 20 seconds

The best China AI model release tracker in 2026 maintains a stable watchlist for 6 core labs — Qwen (Alibaba), DeepSeek, Kimi (Moonshot), GLM (Zhipu), MiniMax, and Step (StepFun) — and uses release cadence, benchmark changes, and license shifts as the three actionable signals. Qwen3 (April 2026) established the current capability ceiling for open-weight Chinese models with the 30B-A3B MoE matching GPT-4o on MMLU under Apache 2.0. DeepSeek R2 is expected H2 2026 based on lab hiring signals. A tracker that covers only headline launches misses the smaller iterative updates (quantized variants, API pricing changes) that often have more immediate builder impact.

Use this page when

  • You need to decide whether a new Chinese model release changes your current stack.
  • You want a weekly routine to track 6 core Chinese AI labs without doomscrolling.
  • You're evaluating Chinese open-weight models for commercial deployment and need license status.

This page is not for

  • Real-time release alerts — use GitHub Watch on individual repos for that.
  • Western model tracking (GPT, Claude, Gemini) — RadarAI has separate pages for those.
  • Academic paper tracking for Chinese AI research — use arXiv or Semantic Scholar.

Key points

  • Qwen3 series (April 2026): 6 models from 0.6B to 235B, all Apache 2.0; Qwen3-30B-A3B (MoE, 3B active params) achieves GPT-4o-class MMLU scores at a fraction of inference cost.
  • DeepSeek V3 (November 2024) remains the most-downloaded Chinese open-weight model on HuggingFace as of May 2026; DeepSeek R2 expected H2 2026.
  • Kimi k1.5 (January 2025) introduced long-context reasoning with 128k window; Moonshot focusing on agentic coding use cases in 2026.
  • GLM-4 (Zhipu, mid-2024) added function-calling parity with GPT-4; GLM-5 expected 2026 with multimodal-first architecture.
  • MiniMax-Text-01 (January 2025) hit 1M token context window — still the longest context of any Chinese model as of May 2026.
  • Track API pricing changes alongside model releases — DeepSeek cut input token prices 3× between V2 and V3 launch.

What changed recently

  • April 2026: Qwen3 launched — 6 model sizes, all Apache 2.0, Qwen3-235B-A22B sets new Chinese open-weight benchmark ceiling.
  • March 2026: DeepSeek confirmed R2 development via researcher social media posts; no official release date.
  • February 2026: GLM-4-Flash API pricing dropped by 60%; Zhipu positioned it as default tier for high-volume enterprise applications.
  • January 2026: MiniMax launched video generation model (Hailuo) as separate product line; text model roadmap shifted to agent-optimized variants.
  • December 2025: Kimi updated agentic coding API with multi-turn tool use; builders report 30–40% reduction in code-writing latency.

Explanation

A model release tracker needs to separate three signal types: (1) capability shifts — when a model meaningfully changes what you can build; (2) pricing/access changes — when inference economics change deployment math; (3) license changes — when open/closed status affects commercial use. Most English media conflates these. Qwen3's Apache 2.0 license was as important as its benchmark scores for builders considering on-premise deployment.

The 6 labs to maintain on a stable watchlist are: Qwen (releases quarterly, Apache 2.0, best open-weight option), DeepSeek (irregular but high-impact releases, MIT license), Kimi/Moonshot (monthly updates, API-first), GLM/Zhipu (enterprise-focused, function-calling strength), MiniMax (long-context specialist), StepFun (multimodal, less frequent but large releases).

Iterative releases (quantized variants, LoRA adapters, API updates) often matter more day-to-day than major version drops. These appear on GitHub and HuggingFace but rarely in English media.

China AI model watchlist — 6 core labs

Track these 6 labs and you cover 90% of the capability surface for Chinese AI models in 2026.

How to verify the answer

Always verify release details from official channels before acting. English media lags 4–6h and often omits technical details like benchmark tables and license terms.

Tools / Examples

  • Qwen3 launch (April 2026): Apache 2.0 across all sizes meant builders could immediately deploy without license review — this was the most actionable part of the announcement for most teams.
  • DeepSeek V3 (Nov 2024): 3× price drop on input tokens at launch made it immediately competitive for high-volume API use cases; this was buried in the pricing table, not the headline.
  • MiniMax-Text-01 1M context (Jan 2025): Builders using long-document workflows switched to MiniMax before switching back to Qwen3 for cost reasons — shows how tracker granularity pays off.

Evidence timeline

Sources

FAQ

Which Chinese AI models should builders track in 2026?

Qwen3 (Alibaba, Apache 2.0, best open-weight), DeepSeek V3/R2 (MIT, cost-efficient API), Kimi k1.5+ (Moonshot, long-context agentic), GLM-4 (Zhipu, function-calling), MiniMax-Text-01 (1M context), Step series (StepFun, multimodal). Track these 6 labs and you cover 90% of the capability surface.

How often do Chinese labs release new models?

Qwen releases major versions roughly quarterly; DeepSeek is irregular (3–4 major releases/year); Kimi updates monthly via API; GLM major versions are semi-annual; MiniMax releases new model families 2–3× per year; StepFun is the least frequent at 1–2 major releases per year.

What is the current capability ceiling for Chinese open-weight models?

As of May 2026: Qwen3-235B-A22B (MoE) sets the ceiling on MMLU, MATH, and coding benchmarks — it matches or exceeds GPT-4o on most published evals. Qwen3-30B-A3B achieves GPT-4o-class performance at significantly lower inference cost due to MoE architecture.

Are Chinese models safe to use commercially?

License varies by model: Qwen3 is Apache 2.0 (permissive commercial use), DeepSeek is MIT, GLM-4 uses a custom license allowing commercial use up to 10M MAU, MiniMax requires commercial license for production. Always check the HuggingFace model card or official repo for current terms.

Where should I check for new Chinese model releases first?

Official lab GitHub repos and HuggingFace pages: QwenLM GitHub (github.com/QwenLM), DeepSeek HuggingFace (huggingface.co/deepseek-ai), Moonshot official blog, Zhipu AI blog. Also @qwen_lm and @deepseek_ai on X/Twitter — these have been 4–6h ahead of English media for major drops.

When should a Chinese model release trigger action vs. just monitoring?

Action triggers: (1) Apache 2.0/MIT license on a model that exceeds your current stack's benchmark threshold; (2) pricing drop that changes the economics of your highest-volume inference calls; (3) context window increase that enables a workflow you've deferred. Monitoring only: incremental benchmark improvements without licensing or pricing changes.

What happened with DeepSeek R2?

As of May 2026, DeepSeek R2 has not been officially announced. Researcher social media posts (March 2026) confirmed active development. Expected H2 2026 based on DeepSeek's historical release cadence (V2: May 2024, V3: November 2024).

How is this page different from China AI company watchlist?

This page tracks model releases — capabilities, benchmarks, licenses, and API changes. The company watchlist tracks which labs and infrastructure companies are worth monitoring at an organizational level. They complement each other: model tracker tells you what shipped, company watchlist tells you who to keep in view.

Search angles this page supports

China AI model release tracker Chinese AI models 2026 Qwen3 release DeepSeek R2 best Chinese AI models track China AI models open weight Chinese models

Related

Go deeper

Last updated: 2026-06-23 · Policy: Editorial standards · Methodology