Moonshot AI dropped the weights for Kimi K3 on July 26 - a day ahead of schedule - and the Hugging Face repo hit 1.56 TB across 96 shards before most teams had finished reading the announcement. Within 48 hours, a handful of hard facts got buried under benchmark charts: the license is not what the previous Kimi model shipped with, the hardware bar is not what "open weight" implies to most people, and the API pricing math contains one number that actually changes the economics.
What Kimi K3 actually is
Kimi K3 is a 2.8T parameter MoE model built on Kimi Delta Attention and Attention Residuals, activating 16 of 896 experts per token - 104B active.
It has a 1,048,576-token context window, four times larger than K2.5, K2.6, and K2.7, and eight times the original K2's 128K.
Artificial Analysis gives K3 a score of 57 on its Intelligence Index and ranks it fourth among 189 models, sitting behind Claude Fable 5 and two GPT-5.6 Sol reasoning settings, then ahead of Claude Opus 4.8, GPT-5.5 at xhigh, and GLM-5.2. The honest read of its coding results is that Moonshot reports narrow K3 wins on Program Bench, SWE Marathon, Automation Bench, BrowseComp, and SpreadsheetBench 2, while the model finishes within half a point of GPT-5.6 Sol on Terminal Bench - but FrontierSWE is a clearer loss: K3 scores 81.2 against Fable 5 at 86.6.
One number worth watching: K3 wins on web-agent tasks with a record 91.2% BrowseComp, costs less than Fable 5, and carries open weights. For teams building browser-based agents or doing document-heavy research workflows, that combination is the actual argument for K3 - not a composite benchmark score.
"Run it locally" means something specific here
This is the part that most launch-day coverage glossed over. Kimi K3 is open-weight, but it is not a desktop model. The live Hugging Face repository occupies 1.561 TB, the official vLLM metadata estimates a 1,680 GB minimum VRAM footprint, and vLLM's deployment recipe calls for at least eight GB300 GPUs.
The naive calculation is 2.8 trillion parameters times 4 bits, which comes to 1.4 TB. The actual repo is 1.56 TB because MXFP4 stores a shared scale factor per block and the non-expert layers sit above 4-bit. Budget 1.56 TB resident just for weights, and 1.8 TB to 2 TB once you add the KV cache and compute buffers. That is the wall, and it does not shrink.
Because K3 is quantization-aware-trained at 4-bit (with 8-bit MXFP8 activations), there is no free halving from the usual 16-bit-to-4-bit step waiting for you; an 8-bit build is bigger than what shipped, not smaller.
In plain hardware tiers:
8× H100 80GB node - minimum for serving; slow generation, reduced context
64+ accelerator supernode - Moonshot's own production guidance; use a high-bandwidth supernode with 64 or more accelerators
Mac Studio, gaming PC, workstation - cannot run K3; a Mac Studio, gaming PC, conventional workstation or small multi-GPU tower cannot do it
The cost math is direct: at roughly $60 an hour before setup overhead, an exploratory weekend on a rented 8-GPU node can cost more than a year of light API use.
If your hardware does not clear the 8× H100 bar, real alternatives exist: OpenRouter, Cursor's Router, the Moonshot API, Together AI, Modal, Fireworks AI, Ollama Cloud, and LM Studio's Secure Cloud all offer hosted access - or self-host Kimi K2.7 Code on far lighter hardware instead.
The license is not what the previous model shipped with
This is the part that matters for any team planning a product. The weights came with a license of their own, no longer based on the previous "Modified MIT" pattern - and one that targets a single group above all: providers who resell the model as a service.
The Kimi K3 License reads like MIT for most of its length, granting anyone the right to use, copy, modify, distribute, sublicense, fine-tune, and sell the weights. One clause then attaches a condition: a licensee running a "Model as a Service" business - defined as giving third parties inference or fine-tuning access with meaningful control over inputs, parameters, or training data - must sign a separate agreement with Moonshot once revenue across the licensee and its affiliates passes $20 million over any consecutive 12 months.
Any commercial product or service with more than 100 million monthly active users, or more than $20 million in monthly revenue, must display "Kimi K3" prominently in its interface.
The sharpest clause in the new license is therefore aimed at the layer above developers: cloud and inference providers that host and resell open models. For the vast majority of users - internal use, research, startups, products with an embedded model - the license behaves in practice like MIT. It becomes relevant for those selling Kimi K3 as a service.
Compare that to K2's actual license text, which only required the attribution display above 100 million MAU or $20 million monthly revenue, with no separate MaaS revenue gate at all. K3 adds a distinct, lower commercial threshold aimed specifically at anyone who wants to resell K3 as an API.
A revenue-conditioned restriction on commercial use is exactly what the Open Source Definition does not allow, so K3's terms came out tighter than K2's, not looser.
If you are building an internal tool, embedding K3 in your own product, or running it for research, the license is effectively permissive. If you are building a paid inference layer on top of it, read the MaaS clause carefully before you ship.
What this means for a real team
Most teams using Kimi K3 should use the API. The Kimi API Platform lists pricing of $3 per million cache-miss input tokens, $0.30 per million cached input tokens, and $15 per million output tokens - making reusable prompts and knowledge bases significantly cheaper. At those rates, a team running a coding agent that processes 10 million input tokens and 2 million output tokens a day spends roughly $60 - the same as one hour of a self-hosted 8-GPU cluster, with nothing operational to maintain.
The case for waiting on self-hosting is also practical. If your project needs open-weight self-hosting, use K2.6 now and migrate to K3 when weights and tooling stabilize - likely late Q3 or early Q4 2026. Community GGUF quantization was not confirmed as of two days post-release, and the vLLM integration was still marked nightly.
A teammate like Beagle doesn't directly run inference, but the same logic applies to any team tool built on a model API: the switch from K2.x to K3 is a config change - provider URL, model name, and a check that your use case clears the license thresholds.
Beagle in action#engineering, 11:02amKimi K3 open weights: common questions
Can I run Kimi K3 on a single machine?
No. The weights are 1.56 TB and require approximately 1,680 GB of VRAM to load. The minimum practical setup is eight H100 80GB GPUs, and Moonshot's own production recommendation is a 64-accelerator supernode. Any machine below that bar - including a well-specced Mac Studio - cannot run K3.
Is Kimi K3 free to use commercially?
For most purposes, yes. The license permits internal use, product embedding, fine-tuning, and distribution without payment. The restriction activates if you run a Model-as-a-Service business and that business generates more than $20 million in revenue over any 12-month period - at that point, a separate agreement with Moonshot is required before commercial use.
How does Kimi K3 compare to Claude Fable 5 on coding?
K3 leads on BrowseComp (91.2%), Program Bench, and SWE Marathon in Moonshot's own table. Fable 5 leads on FrontierSWE (86.6 vs 81.2) and the Artificial Analysis long-horizon coding tracker. The differences on most benchmarks are under one point, which means harness choice and retry policy matter more than the raw number.
What is the Kimi K3 API price?
$3 per million input tokens and $15 per million output tokens via the Moonshot platform and OpenRouter. Cached inputs drop to $0.30 per million tokens, which significantly changes economics for agents that repeatedly reference the same large context.
Should a team self-host Kimi K3 or use the API?
Use the API unless you have a specific privacy, compliance, or latency requirement that demands on-premises weights. Self-hosting K3 costs roughly $60 per hour on a rented 8-GPU cluster, before setup overhead - more than a year of moderate API use for most teams. Wait for tooling to stabilize and community GGUF quantization to land, then revisit in Q4 2026.