{
  "name": "China Frontier LLM Release Tracker",
  "url": "https://chinaaidispatch.com/trackers/frontier-models/",
  "publisher": "China AI Dispatch",
  "datePublished": "2026-07-17",
  "dateModified": "2026-07-28",
  "asOf": "2026-07-28",
  "description": "A sourced release tracker of China’s frontier large language models: model, lab, release date, parameters and context, license, public pricing and a benchmark claim, each row attributed to a named source.",
  "note": "Benchmark claims are marked as lab-reported or independent; vendor benchmarks are not presented as neutral fact. Undisclosed specs are left empty rather than estimated.",
  "rowCount": 9,
  "rows": [
    {
      "id": "deepseek-v4-pro",
      "model": "DeepSeek V4-Pro",
      "lab": "DeepSeek",
      "released": "24 Apr 2026",
      "releasedIso": "2026-04-24",
      "params": "1.6T total, 49B active (MoE)",
      "context": "1M tokens",
      "license": "Open-weight (MIT)",
      "pricing": "About $3.48 per million output tokens, roughly one-seventh the price of Claude Opus 4.7",
      "benchmark": "Coding a hair behind the best Western models; the DeepSeek report’s 80.6% on SWE-Bench Verified was independently confirmed, behind Opus 4.7 at 87.6% but leading on LiveCodeBench",
      "benchmarkReportedBy": "independent",
      "source": "DeepSeek technical report; Reuters; independent benchmarks",
      "url": "https://chinaaidispatch.com/trackers/frontier-models/#deepseek-v4-pro",
      "companyUrl": "https://chinaaidispatch.com/companies/deepseek/",
      "cadIssueUrl": "https://chinaaidispatch.substack.com/p/the-week-the-race-happened"
    },
    {
      "id": "qwen-3-7-max",
      "model": "Qwen3.7-Max",
      "lab": "Alibaba (Tongyi/Qwen)",
      "released": "20 May 2026",
      "releasedIso": "2026-05-20",
      "params": "Not disclosed (Alibaba does not publish Qwen parameter counts)",
      "license": "Proprietary (API only)",
      "pricing": "Per-token price not disclosed by Alibaba",
      "benchmark": "Highest-ranked Chinese model on the LMArena leaderboard, 13th globally, between GPT-5.5 and Grok 4.2; fifth globally on Artificial Analysis",
      "benchmarkReportedBy": "independent",
      "source": "LMArena; Artificial Analysis; Alibaba Qwen team; InfoQ",
      "url": "https://chinaaidispatch.com/trackers/frontier-models/#qwen-3-7-max",
      "cadIssueUrl": "https://chinaaidispatch.substack.com/p/the-stack-completes"
    },
    {
      "id": "glm-5-2",
      "model": "GLM-5.2",
      "lab": "Zhipu (Z.ai)",
      "released": "13 Jun 2026 (API and MIT open-source rolled out the following week)",
      "releasedIso": "2026-06-13",
      "params": "Not disclosed at release",
      "context": "1M tokens",
      "license": "Open-weight (MIT)",
      "pricing": "Roughly one-sixth the API cost of GPT-5.5 (VentureBeat). On the Harvey legal benchmark it averaged about $2.40 per task against about $31 for Anthropic’s Fable, a roughly 13x spread (Applied Compute case study, via CAD, 28 Jul 2026). Absolute per-token price not disclosed",
      "benchmark": "First open-weight model to land at 51 on the Artificial Analysis Intelligence Index, between GPT-5.5 and Opus 4.8; 74.4 on FrontierSWE, within a point of Opus 4.8",
      "benchmarkReportedBy": "independent",
      "source": "SCMP; VentureBeat; Artificial Analysis; Applied Compute; tmtpost",
      "url": "https://chinaaidispatch.com/trackers/frontier-models/#glm-5-2",
      "cadIssueUrl": "https://chinaaidispatch.substack.com/p/the-daily-driver"
    },
    {
      "id": "glm-5-1",
      "model": "GLM-5.1",
      "lab": "Zhipu (Z.ai)",
      "released": "8 Apr 2026",
      "releasedIso": "2026-04-08",
      "params": "754B (MoE) per post-launch coverage; a pre-release third-party estimate cited 744B",
      "license": "Open-weight (MIT)",
      "pricing": "Priced close to Claude Sonnet after a launch increase; absolute per-token price not disclosed",
      "benchmark": "Zhipu claims #1 globally on SWE-bench Pro, above GPT-5.4 and Opus 4.6; independent breakdowns put it at 94.6% of Opus 4.6",
      "benchmarkReportedBy": "lab",
      "source": "Zhipu launch docs; WaveSpeed; tmtpost; Reuters",
      "url": "https://chinaaidispatch.com/trackers/frontier-models/#glm-5-1",
      "cadIssueUrl": "https://chinaaidispatch.substack.com/p/the-price-war-is-over"
    },
    {
      "id": "kimi-k3",
      "model": "Kimi K3",
      "lab": "Moonshot AI",
      "released": "Week of 17 Jul 2026",
      "releasedIso": "2026-07-17",
      "params": "2.8T per Qbit’s launch-day report (called the largest open-weight model to date); pre-release estimates ranged 2.5T to 3T",
      "license": "Open-weight",
      "benchmark": "Took the top spot on the Frontend Code Arena, ahead of the leading US models; Chinese models still trail the top US labs on the hardest independent coding benchmarks",
      "benchmarkReportedBy": "independent",
      "source": "Qbit (量子位); Sina Finance; 36Kr",
      "url": "https://chinaaidispatch.com/trackers/frontier-models/#kimi-k3",
      "cadIssueUrl": "https://chinaaidispatch.substack.com/p/the-cheaper-teacher"
    },
    {
      "id": "kimi-k2-6",
      "model": "Kimi K2.6",
      "lab": "Moonshot AI",
      "released": "20 Apr 2026",
      "releasedIso": "2026-04-20",
      "params": "1T (MoE)",
      "license": "Modified MIT",
      "pricing": "$0.95 in / $0.16 cached / $4.00 out per million tokens; API prices raised 58% on release",
      "benchmark": "Moonshot claims 58.6% on SWE-Bench Pro, beating every closed-source model, and 54.0% on Humanity’s Last Exam with tools; independently placed at 54 on the Artificial Analysis index, tied with DeepSeek V4 and Qwen",
      "benchmarkReportedBy": "lab",
      "source": "Moonshot benchmark; The Batch (deeplearning.ai); Artificial Analysis; 36Kr",
      "url": "https://chinaaidispatch.com/trackers/frontier-models/#kimi-k2-6",
      "cadIssueUrl": "https://chinaaidispatch.substack.com/p/300-agents-and-a-price-hike"
    },
    {
      "id": "minimax-m3",
      "model": "MiniMax M3",
      "lab": "MiniMax",
      "released": "1 Jun 2026",
      "releasedIso": "2026-06-01",
      "params": "Not disclosed",
      "context": "1M tokens",
      "license": "Open-source (weights released about 10 days after the 1 Jun launch)",
      "pricing": "¥119 per month for 18 billion tokens, roughly 15x the token volume of a comparable Claude plan",
      "benchmark": "Global rank #7 on the Artificial Analysis intelligence index; MiniMax reports 93.2% on GPQA Diamond, above Claude Opus 4.8 and 4.7",
      "benchmarkReportedBy": "independent",
      "source": "Artificial Analysis; MiniMax technical report",
      "url": "https://chinaaidispatch.com/trackers/frontier-models/#minimax-m3",
      "cadIssueUrl": "https://chinaaidispatch.substack.com/p/the-leaderboard"
    },
    {
      "id": "tencent-hunyuan-hy3",
      "model": "Hunyuan Hy3",
      "lab": "Tencent Hunyuan",
      "released": "Apr 2026 preview (open-sourced Jul 2026)",
      "releasedIso": "2026-04-23",
      "params": "295B total, 21B active (MoE)",
      "context": "256K tokens",
      "license": "Open-weight (open-sourced after an open preview)",
      "pricing": "About 1.2 yuan ($0.17) per million input tokens on Tencent Cloud, 0.4 yuan on a cache hit",
      "benchmark": "Topped OpenRouter’s weekly token-usage chart at 3.66 trillion tokens (a usage metric, not a quality test); InfoQ rated it top-tier on coding-agent and instruction-following tasks among comparably sized models at launch",
      "benchmarkReportedBy": "independent",
      "source": "OpenRouter; tmtpost; InfoQ; Reuters",
      "url": "https://chinaaidispatch.com/trackers/frontier-models/#tencent-hunyuan-hy3",
      "cadIssueUrl": "https://chinaaidispatch.substack.com/p/88-days"
    },
    {
      "id": "meituan-longcat-2",
      "model": "LongCat-2.0",
      "lab": "Meituan (LongCat)",
      "released": "30 Jun 2026 (open-sourced at release)",
      "releasedIso": "2026-06-30",
      "params": "1.6T total, about 48B active (MoE)",
      "context": "1M tokens",
      "license": "Open-weight",
      "benchmark": "Meituan reports it as the first trillion-parameter model trained and inferenced entirely on a 50,000-card domestic cluster with no foreign accelerators, on over 30 trillion tokens",
      "benchmarkReportedBy": "lab",
      "source": "SCMP; Reuters; Geek Park; InfoQ",
      "url": "https://chinaaidispatch.com/trackers/frontier-models/#meituan-longcat-2",
      "cadIssueUrl": "https://chinaaidispatch.substack.com/p/the-harder-half"
    }
  ]
}