Not proven yet, but Harvey is building on it. In August 2026, Harvey, an OpenAI-backed legal AI company, announced Tenet, a research-preview legal model post-trained on Kimi K3. AWS added K3 to Amazon Bedrock in September, and Microsoft was reportedly testing it in July. For Australian teams, the next question is who runs inference, and where.
What did Harvey build on Kimi K3?
Harvey built Tenet, a legal model post-trained on a Kimi K3 base. In a 20 August 2026 post, Harvey called it its first post-trained open-weight model. Harvey post-trained it with a research partner for long-horizon legal work.
Harvey reports that Tenet completes almost twice as many held-out tasks as base Kimi K3 on LAB, its Legal Agent Benchmark, and 20% more on LAB contracts tasks. Harvey puts the gains at 9 and 2 percentage points of all-pass rate. These are company-reported results, and Harvey labels Tenet a research preview.
Harvey states that it did not use any customer data in its post-training.
Why would an OpenAI-backed company choose an open-weight base?
Harvey's post says its research over the past six months had two goals. The first is building frontier legal intelligence on open-weight models. The second is letting law firms build their own specialised models and own their intelligence.
Published weights make that second goal practical. A team can post-train them and choose where the model runs. That is harder with a model you can only reach through the vendor's API.
It is a change of direction for Harvey. The Star, reporting SCMP on 29 August 2026, said the start-up had focused on customising closed models from Anthropic, OpenAI and Google. The Next Web described the move as Harvey no longer renting the thing its business depends on.
For the wider trade-offs, see the guide to open-weight AI models in 2026.
Who else offers or tests Kimi K3?
AWS offers it on Bedrock, and Microsoft was reportedly testing it in July. Moonshot AI released Kimi K3 on 16 July 2026, according to The Deep Dive.
On 18 September 2026, AWS announced Kimi K3 on Amazon Bedrock. AWS says it is the first open-weight model on Bedrock to support explicit prompt caching. AWS also cites Moonshot AI's own figures. Moonshot calls K3 its most capable model, with 2.8 trillion parameters and a 1-million-token context window.
The Microsoft news is less certain. The Deep Dive reported on 21 July 2026, citing The Information, that Microsoft was testing K3 on Azure for Copilot tasks. The report said Microsoft engineers must first confirm that K3 meets Copilot's requirements for quality, reliability, safety and latency.
Microsoft had not named which Copilot tasks might move, so treat this as reported testing, not a rollout. The same report noted that GitHub Copilot had offered an earlier model, Kimi K2.7 Code, since 1 July 2026.
Does a model's origin decide where your data goes?
Not on its own. For data location, the provider that runs inference matters more than where the model came from.
The Next Web noted that Harvey had put open weights underneath privileged client work. A host answers part of that concern, because it decides where prompts are processed. Licence terms and model provenance still sit with the model, so check those separately.
Routing choices matter even within one host. At launch on 18 September 2026, AWS offered K3 on Bedrock through two cross-Region inference profiles. AWS said the global profile routes each request to any supported commercial AWS Region worldwide. It said the US profile keeps processing within the US geography.
AWS also states that Bedrock processes data within the AWS data boundary. It says the data is not shared with the model provider or used to train the underlying model. That is AWS's own claim, as of September 2026.
The guide to data sovereignty and AI in Australia covers what to ask a host in writing.
Where can you run Kimi K3?
You have four broad options. The table compares them on the points a data owner asks about first.
| Option | Who runs inference | Where requests are processed | What you manage |
|---|---|---|---|
| The model vendor's own API | The model vendor | Set by the vendor, so check its terms | An API key and usage |
| A big-cloud platform, such as Bedrock | The cloud provider | The regions the provider offers for that model, so check its current list | Cloud account, permissions and routing choice |
| An API hosted in Australia | The API provider | Ask the provider to confirm it in writing | An API key and usage |
| Self-hosting the weights on your own or dedicated GPUs | You, or your infrastructure team | Wherever your GPUs sit | GPUs, serving engine, scaling and updates |
The right choice depends on what you already run. A team already on a big cloud may start there if that cloud's routing options suit it. A team that wants to own a specialised model, as Harvey describes, may look at self-hosting.
Whatever you pick, ask each provider the same questions in writing. Where are requests processed? Are prompts or outputs used for training? Can you change the routing? How does the provider handle new model versions?
How Xinference fits
Xinference Model API serves Kimi K3 and DeepSeek V4.1 Flash, hosted in Australia on dedicated GPUs, at the same per-token price as the model vendor. Kimi K3 is $3.00 in and $15.00 out per 1M tokens, in USD (Kimi K3 API). See pricing.
No training on your data. No prompt or response content kept by default.
It uses an OpenAI-compatible API, so moving from another OpenAI-compatible provider is a base URL and API key change. Moving later to dedicated GPUs or servers you run yourself is also a base URL change.
Don't see the model you need? Talk to us about adding it. Read more on the Model API page, or Get API access.
Questions
Is Harvey Tenet available to use?
Harvey published Tenet as a research preview on 20 August 2026. Its post does not describe a general release, so ask Harvey about access.
Did Harvey train Tenet on customer data?
No, according to Harvey. Its August 2026 post states that it did not use any customer data in any of its post-training efforts.
Is Kimi K3 open source?
Kimi K3 is open-weight, which means the trained weights are published so others can run and post-train them. "Open-weight" describes access to the weights, not the licence terms. Read Moonshot AI's licence before commercial use.
How much does Kimi K3 cost on Xinference?
On the Xinference Model API, Kimi K3 is $3.00 in and $15.00 out per 1M tokens, in USD, as of October 2026. That is the same per-token price as the model vendor. See pricing.