Kimi K3 is open. The easy part ends there.
When I first read “open weights”, I assumed this was mainly an access story. Then I reached Moonshot’s deployment recommendation: 64 or more accelerators. That one detail changes the whole decision.
Most teams should not begin by downloading 1.56TB of model files. They should begin with one hosted or API test and ask whether Kimi K3 handles a real job well enough to justify changing anything.
- The Big Signal: Why 64 accelerators matter more than another benchmark win
- Build This: Choose hosted, API or self-hosted access with one repeatable test
- Worth Watching: Check the software people already use before adding another AI layer
Kimi K3 widens model choice without making deployment simple
My view: the open weights matter, but the practical opportunity is to test the model without inheriting a small infrastructure department.
Kimi K3 was already available through Moonshot’s products and API. The public repository now adds the full model weights, code and licence, giving technical teams a third route: run the model on infrastructure they control.
- Scale: 2.8 trillion total parameters, with roughly 104 billion activated for each token.
- Capability: Text and image input with a context window of 1,048,576 tokens.
- API price: $0.30 per million cached input tokens, $3 per million uncached input tokens and $15 per million output tokens.
- Deployment: The Hugging Face repository is approximately 1.56TB, and Moonshot recommends supernodes with 64 or more accelerators.
- Independent evidence: Artificial Analysis ranked Kimi K3 second on AA-Briefcase, but it averaged $10.57 and 56.4 minutes per task.
- Hosted product Quickest route for occasional research, coding and document work.
- API Best starting point for repeatable tests, cost tracking and controlled integrations.
- Self-hosted weights Most control, with the hardware, security and maintenance bill attached.
How to try it: Use Kimi.com, Kimi Work, Kimi Code or the kimi-k3 API. Start with a public report-to-spreadsheet task containing five figures you can check yourself.
Status: Open weights · API available · Not tested by Future Relay
The detail I would put in red
The leaderboard results are interesting. The 64-accelerator recommendation is the part that changes a business decision. Self-hosting Kimi K3 is less like installing a new app and more like deciding to operate a small slice of a cloud provider.
The licence deserves the same care. It permits broad use, modification and distribution, but includes separate conditions for large model-as-a-service businesses and prominent naming requirements for very large commercial products. “Open weights” describes access to the model; it does not remove legal, operational or security work.
Open models expand who can inspect, adapt and host frontier AI. That is valuable. It still does not answer the only question that matters to most teams: what becomes materially better if we take responsibility for running it?
Your move
Test one difficult but reversible task against your current model. Record the finished quality, corrections, elapsed time, cost and data concerns. Move beyond the hosted product only when the result gives you a specific reason to do so.
Explore the Kimi K3 model and weightsIndependent evidence: read the AA-Briefcase evaluation
Create a Model Access Decision Card
The outcome is a one-page record showing whether a model belongs in a hosted product, an API workflow or infrastructure you operate yourself.
You’ll need
- Kimi K3 through its hosted product or API
- Your current comparison model
- A spreadsheet or document
Five steps
- Choose one reversible task. Use public or synthetic information, such as turning a public report into a small spreadsheet or revising a non-production code sample.
- Define success first. Record the required deliverable, factual checks, acceptable completion time and maximum cost before either model begins.
- Run the same task twice. Keep the input, tools and completion criteria consistent across Kimi K3 and your current model.
- Judge the finished work. Check facts, calculations, citations, missing requirements, human corrections, latency and unexpected actions—not just the first impressive paragraph.
- Select the access route. Use the hosted product for occasional work, the API for repeatable workflows and self-hosting only when control, volume or privacy genuinely pays for the burden.
Task tested: [ ]
Success criteria: [ ]
Data sensitivity: Public / Internal / Confidential
Required deliverable: [ ]
Factual accuracy: [ ]
Requirements completed: [ ]
Human corrections: [ ]
Completion time: [ ]
Estimated cost: [ ]
Integration effort: [ ]
Privacy or jurisdiction concern: [ ]
Hosted-product result: [ ]
API result: [ ]
Self-hosting justification: [ ]
Decision: Adopt / Pilot / Watch / Reject
Review date: [ ]
Pro tip
A model can win a benchmark and still lose your workflow. Do not upload confidential, personal or regulated information during the first test, and record the awkward failures as carefully as the polished successes.
Read Kimi’s API data and security guidanceGoogle Classroom puts useful information at the front door
This is not the week’s most glamorous update. It may be the one that prevents more duplicated work.
Google began rolling out a role-based Classroom homepage on 27 July. Teachers receive assignment and student-interaction information, students see work due soon, and school leaders gain high-level analytics and administrative shortcuts.
The rollout covers Workspace customers, Workspace Individual subscribers and personal Google accounts. The exact modules depend on the person’s role, account and enabled features, and there is no switch to keep the previous homepage.
The broader lesson is easy to miss: before adding another dashboard, reminder bot or AI summary, check whether the software people already open each morning has quietly moved the information to the front door.
See Google’s Classroom rollout detailsNVIDIA and more than 40 founding partners have formed the Open Secure AI Alliance to develop shared models, harnesses and security tools. The open-source NOOA research framework is already available for testing, tracing, auditing and governing agent behaviour; the wider alliance is still at the building stage. Status: Alliance launched · Initial framework available.
Explore the Open Secure AI AllianceI would begin with the API and one awkward, verifiable task—not the biggest task and not the safest demo. Reply with hosted app, API, self-hosted or not yet, and tell me what you would test first.
Until next time,
Tom
Future Relay
