Harvey: 550 million dollars and models on Chinese bases

{
  "signal_id": "cc397c0f-1873-544a-a07e-1069c9522f05",
  "signal_type": "news",
  "signal_subtype": "is_developing",
  "detected_at": "2026-09-10T08:00:56+00:00",
  "company": {
    "name": "Harvey",
    "domain": "harvey.ai"
  },
  "data": {
    "url": "https://fomoera.com/harvey-levanta-550-mdd-para-dar-modelo-propio-a-los-despachos-los-suyos-parten-de-bases-chinas/",
    "title": "Harvey: 550 million dollars and models on Chinese bases - fomoera.com",
    "excerpt": "Harvey, the artificial intelligence company for law firms, announced on September 9, 2026, a 550 million dollar round at a valuation of 15.5 billion dollars, co-led by Diffusion and Lightspeed Venture Partners. Its statement says the money will accelerate the company's work to help law firms and in-house legal departments build and \"own their intelligence\" at scale. The technical report it published on August 20 explains what that intelligence is made of today: its first proprietary model, Harvey Tenet, is based on Kimi K3, the open weights from China's Moonshot AI, and the other three capabilities it trained come from GLM-5.2, from Z.ai, and from Qwen3.8-27B, from Alibaba. None of these bases were built by Harvey.\nHarvey took Kimi K3, a mixture-of-experts model with 2.8 trillion parameters that Moonshot AI released in July 2026, and post-trained it together with Fireworks using asynchronous reinforcement learning on about 1,750 legal task environments. Each environment includes a client's file, an instruction drafted as a partner's assignment, and a rubric that breaks down that partner's review into pass or fail criteria. The training ran on about 150 NVIDIA B300 GPUs for two months, and the company clarifies that it did not use client data anywhere in the process.\nTenet completes almost double the reserved LAB tasks compared to the Kimi K3 base and 20% more in LAB Contracts...",
    "product": "Harvey Tenet",
    "summary": "Harvey is developing its first proprietary model, Harvey Tenet, based on open weights from other AI labs.",
    "planning": true,
    "image_url": "https://fomoera.com/content/images/2026/09/fomoera-harvey-levanta-550-mdd-para-dar-modelo-propio-a-los-despachos-los-suyos-parten-de-bases-chinas-74ba6b6b-16x9-v2-1600x900.webp",
    "confidence": 0.9,
    "product_data": {
      "name": "Harvey Tenet",
      "full_text": "its first proprietary model, Harvey Tenet",
      "fuzzy_match": false,
      "release_type": "model"
    },
    "product_tags": [
      "future_tech",
      "general_technology"
    ],
    "published_at": "2026-09-10T08:00:56Z",
    "article_sentence": "The technical report it published on August 20 explains what that intelligence is made of today: its first proprietary model, Harvey Tenet, is based on Kimi K3, the open weights from China's Moonshot AI, and the other three capabilities it trained come from GLM-5.2, from Z.ai, and from Qwen3.8-27B, from Alibaba."
  }
}

curl https://signals.autobound.ai/v1/signals/cc397c0f-1873-544a-a07e-1069c9522f05 \
  -H "X-API-KEY: $AUTOBOUND_API_KEY"

curl -X POST https://signals.autobound.ai/v1/companies/enrich \
  -H "X-API-KEY: $AUTOBOUND_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"domain":"harvey.ai","limit":20}'