What happened
The Star, republishing a South China Morning Post report, says Harvey built and released Harvey Tenet on top of Moonshot AI’s Chinese Kimi K3 open-weight model. Harvey says the system was trained for complex legal work and outperformed its base model and several US frontier systems in internal evaluations.
The Star reports that San Francisco-based legal technology company Harvey released Harvey Tenet on Thursday after building the system on top of Moonshot AI’s Kimi K3, a Chinese open-weight model. Harvey said Tenet was post-trained for complex legal work. The company described the release as the result of six months of research into using open-weight models to create what it called “frontier legal intelligence” and to help law firms build and deploy specialized models. The reported account thus presents Tenet as a specialized legal system built through adaptation of an existing open-weight foundation. It describes the research and release together, while leaving the product’s broader deployment details for follow-up.
According to The Star’s report, Harvey previously concentrated on customizing closed proprietary models from Anthropic, OpenAI and Google for legal applications. The company now says it trained Tenet on comprehensive legal data sets and that the model outperformed both Kimi K3 and US frontier systems identified in the article as Fable 5 and GPT-5.6 Sol across a range of complex, long-horizon legal agentic tasks. Those results are company claims; the report does not provide the underlying test set, scores, methodology or independent replication. The report therefore gives a description of Harvey’s own comparison, rather than a complete account of how the systems were tested. The absence of scores and methodology makes the relative performance difficult to assess from the report alone.
Harvey said the training process took two months and used about 150 Nvidia B300 graphics processing units. The company also said it sought to improve cost efficiency in two ways: by starting with an open-weight model with lower per-token prices and by reducing the number of tokens consumed during inference. The report places the announcement in the context of growing US enterprise interest in open models, while citing a separate Information report that open-source models account for 40% of AT&T employee AI queries. The article says AT&T is not currently using Chinese open-weight models and is still evaluating risks associated with options including DeepSeek and Kimi K3. The report does not quantify the resulting savings, so the cost-efficiency benefit remains a stated objective rather than a measured result in the available account. Its AT&T comparison supplies context but does not establish use of Kimi K3 there.
Source details: thestar.com.my ↗
Why it matters
The move is a concrete example of a Western enterprise AI company using a Chinese open-weight model as the foundation for a specialized commercial system. It also illustrates how post-training, rather than building a general-purpose model from scratch, may let companies customize AI for particular industries while seeking lower operating costs.
The reported release matters because it puts an open-weight Chinese model at the center of a commercial legal product backed by prominent US investors and used by major international law firms and enterprise clients. The development is therefore more consequential than a routine model experiment: it concerns the foundation of a specialized system intended for professional work where accuracy, confidentiality and reliability are central requirements.
Harvey’s approach shows one route for enterprise AI development. Instead of training a general-purpose model from the beginning, a company can start with an existing base model and apply post-training to industry-specific data and tasks. In principle, that can give the company more control over specialization and deployment while lowering some model-development and inference costs. The practical value of Harvey Tenet will depend on whether the reported gains hold on representative legal work and whether the system can meet the quality and confidentiality requirements of law firms.
The announcement also highlights a geopolitical and commercial tension. The report describes Chinese open-weight models as increasingly relevant to Western developers, while noting that US companies are still assessing the risks of using them. Open weights may offer greater technical flexibility, but the source does not establish how Kimi K3 is licensed, where Tenet is hosted, what safeguards Harvey applies, or how the company addresses data-provenance and security concerns. Those unanswered questions are especially important for legal applications involving sensitive client information.
What to watch next
Independent testing will be important because the reported performance comparisons come from Harvey. Watch for details about Harvey Tenet’s benchmarks, customer availability, model licensing, data governance, security review, and whether other Western companies follow a similar path with Chinese open-weight systems.
The first priority is independent evaluation. Harvey’s claims about state-of-the-art performance and superiority to Kimi K3, Fable 5 and GPT-5.6 Sol are not independently confirmed in the source. Useful follow-up reporting would identify the legal tasks tested, the comparison systems’ configurations, the number and type of cases, the error rates, the role of human review and whether results reproduce outside Harvey’s own evaluation environment.
Availability and governance will also determine the product’s significance. The report does not say whether Harvey Tenet’s weights, training methods or evaluation materials will be publicly released, or whether the product is generally available to customers. It also does not specify the licensing terms for Kimi K3, the jurisdictions in which Tenet can be deployed, how customer data is isolated, or what retention and audit controls apply to legal documents and prompts.
Finally, watch whether this remains a company-specific choice or becomes a broader enterprise pattern. The Star cites AT&T’s reported use of open-source models but says the company is not using Chinese open-weight models at present and is evaluating their risks. Further deployments, security reviews, procurement policies or restrictions would clarify whether Chinese open-weight systems can gain practical acceptance in sensitive Western industries. The source provides no evidence yet about wider adoption, customer outcomes or regulatory responses.