Kimi K3: what a frontier-class open model changes for your sovereignty and your bill

Transparency note. This article follows the IgnitionAI editorial policy. The model specifications and licence clauses come from the official Hugging Face model card, Moonshot AI's technical blog and the original LICENSE file, accessed on 2 August 2026. Updated on 24 October 2026 for Kimi K3.1 Pro, the actual checkpoint size and single-node hosting (Moonshot AI blog, Artificial Analysis, Hugging Face API, vLLM recipe, accessed on 24 October 2026). The hosting cost scenarios rely on prices collected on 24 October 2026, in a very volatile GPU hardware market. Cost and sizing estimates are tagged as IgnitionAI estimates. IgnitionAI has no commercial relationship with Moonshot AI.

Moonshot AI released the Kimi K3 weights on 27 July 2026: 2.8 trillion parameters, 104 billion of them activated per token, and a one-million-token context window. For most mid-market companies, this release changes nothing about whether you can host the model: the weights alone need about 1.4 TB by calculation, 1.56 TB published, in GPU memory. It changes three other things, which matter more: your reversibility, the market's price floor, and the feasibility of sovereign hosting.

And its licence is not an open-source licence. That is the part almost nobody reads.

What Kimi K3 is, factually

Specification	Value
Total parameters	2,800 billion (2.8T)
Parameters activated per token	104 billion, that is 16 experts out of 896
Context window	1,048,576 tokens
Modalities	text and image, natively
Quantisation	MXFP4 weights, MXFP8 activations, via quantisation-aware training
Weights released	27 July 2026, on Hugging Face
Licence	Kimi K3 License (not an open-source licence, see below)

Source: Hugging Face model card and Moonshot AI technical blog, accessed 2 August 2026.

The architecture rests on two mechanisms Moonshot calls Kimi Delta Attention and Attention Residuals, paired with a mixture-of-experts framework named Stable LatentMoE. The vendor claims roughly 2.5 times better scaling efficiency than Kimi K2, and presents the model as the first 3-trillion-class model published with open weights.

Sparsity is the point to retain as a decision-maker. The model holds 2.8 trillion parameters but activates only 104 billion per token. Compute cost per request comes closer to that of a 100-billion-parameter model, while the memory required stays that of a 2.8-trillion-parameter one. This split between compute cost and memory cost is exactly what makes the model attractive to serve, and hard to host.

Performance, without the hype

The results Moonshot publishes on its model card:

Evaluation	Score
GPQA Diamond (scientific reasoning)	93.5
Terminal-Bench 2.1 (terminal tool use)	88.3
BrowseComp (agentic web research)	91.2
DeepSWE (software engineering)	67.5
Video-MME (video understanding)	90.0
MathVision (visual mathematics)	94.3 without tools, 97.8 with tools

Source: Hugging Face model card, accessed 2 August 2026. Scores self-reported by the vendor.

Two reading precautions, which apply to every model.

First, these scores are self-reported. They are useful to place an order of magnitude, not to settle an architecture choice. Only an evaluation on your own data and your own queries will tell you what the model is worth for your case.

Second, the protocol matters as much as the score. Moonshot notes that the BrowseComp score of 91.2 is obtained with context compaction at 300,000 tokens, and that the same evaluation run on the full one-million-token window without compaction gives 90.4. The gap is small, but it illustrates a general rule: a benchmark without its protocol is not usable data.

On positioning, the vendor is explicit itself. Moonshot's technical blog states that K3's overall performance still trails the most powerful proprietary models, which it names as Claude Fable 5 and GPT-5.6 Sol, while claiming frontier-level performance across its evaluation suite. That candour from the vendor is a sign of seriousness, and it saves you from basing a decision on a marketing comparison.

Update, 24 October 2026. This holds for K3. On 23 October 2026, Moonshot AI launched Kimi K3.1 Pro, a version of K3 post-trained with much more compute, available through the Kimi API and apps. Artificial Analysis gives it 55 points on its Intelligence Index, against 44 for K3. It ranks as the 4th model, behind Claude Opus 5.5 (58), Claude Sonnet 5.5 and Muse Wonder Full (56). Its weights are announced by Moonshot for 4 November 2026 at the latest, but they are not published yet, and its licence is not known yet. The licence and hosting figures below therefore refer to K3.

The trap nobody reads: the licence

"Open weights" does not mean "open source". Kimi K2 shipped under a modified MIT licence. K3 ships under a new document, the Kimi K3 License, which introduces commercial thresholds.

What the licence permits, in its terms: use, copy, modify, merge, publish, distribute, sublicense and sell copies of the software, as well as run, deploy, fine-tune it and create derivative works from it.

What it conditions:

• Model-as-a-service threshold. If you or your affiliates operate a model-as-a-service business exceeding 20 million US dollars over any consecutive twelve months, a separate agreement with Moonshot AI is required before commercial use.
• Attribution requirement. If the software powers a commercial product exceeding 100 million monthly active users, or 20 million US dollars in monthly revenue, "Kimi K3" must be prominently displayed in the user interface.
• Exemptions. These requirements apply neither to internal use not made available to third parties, nor to use through Moonshot AI's official products or certified partners.

Source: official LICENSE file, accessed 2 August 2026. The software is provided "as is", without warranty.

Translation for three profiles. A mid-market company deploying the model for its own staff, without reselling it, falls under the internal-use exemption: none of these constraints apply. A consultancy or vendor reselling inference as a service must watch the 20-million-dollar threshold. A high-audience consumer product must plan for the interface attribution.

The lesson goes beyond Kimi K3: an open model's licence is a contractual clause to read at scoping, on a par with a supplier contract. Open-model licences increasingly diverge, and a "free" model can carry obligations that surface exactly when your product takes off.

The calculation everyone skips: hosting 2.8 trillion parameters

Let's do the arithmetic. At the model's native quantisation, four bits per parameter, 2.8 trillion parameters occupy roughly 1.4 TB on paper. The checkpoint published on Hugging Face actually weighs 1.56 TB, because part of the weights stays in 16 bits. That is the memory for the weights alone, before the attention cache, which grows with context length and concurrent users.

A single server can serve K3, but not just any server. You need a node of 8 GPUs with 288 GB each, such as NVIDIA B300 or AMD MI355X, about 2.3 TB in total: the vLLM recipe serves K3 on one node of 8 MI355X, and a deployment on 8 B300 is documented. Nodes of 8 H200 (141 GB per GPU), 8 B200 (180 GB) or 8 MI300X (192 GB) are not enough, even for the weights alone. IgnitionAI estimate: such infrastructure, rented from a cloud provider, runs into tens of thousands of euros per month, before the operations skills needed to keep it running.

By comparison, API access to K3 was priced at around 2.90 dollars per million input tokens and 14 dollars per million output tokens (OpenRouter pricing, accessed 2 August 2026). Kimi K3.1 Pro costs 2 dollars per million input tokens and 10 dollars per million output tokens (Artificial Analysis and OpenRouter, accessed 24 October 2026).

The break-even point is simple to compute:

monthly break-even volume ≈ monthly infrastructure cost / average API price per token

IgnitionAI estimate: for the vast majority of mid-market companies, whose consumption runs to tens or hundreds of millions of tokens per month, the API stays far cheaper than dedicated hosting. Self-hosting a model of this size is justified by sovereignty or by industrial volume, rarely by economics alone.

Put differently: publishing the weights does not mean you will host this model. It means someone can host it for you, without depending on the original vendor. That is where the real change lies.

Rent, buy or stay on the API: priced scenarios as of 24 October 2026

Prices collected on 24 October 2026. GPU hardware prices are moving sharply right now, for rental and purchase alike. The figures in this section come from a single day. Redo the calculation with your own quotes before any decision.

Does hosting K3 yourself pay off, and from what volume? Answering without guessing takes four numbers: the API price, the cost of an 8 × B300 node rented or bought, its power draw, and how many tokens it serves. Here they are, each with its source.

Prices collected

Item	Value collected	Source, accessed on 24 October 2026
Kimi K3.1 Pro API	2 dollars per million input tokens, 10 dollars output, 0.20 dollars cached	OpenRouter, Artificial Analysis
Kimi K3 API, price displayed by OpenRouter	0.239 dollars input, 10.80 dollars output	OpenRouter, Kimi K3
Kimi K3 API, served by Moonshot AI via OpenRouter	3 dollars input, 15 dollars output	OpenRouter, K3 providers
8 × B300 rental, Massed Compute	52.80 dollars per hour	massedcompute.com/pricing
8 × B300 rental, Scaleway, Paris (fr-par-2)	1 euro per minute excluding VAT, i.e. 60 euros per hour	Scaleway public catalogue, GPU pricing page
HGX B300 rental, Nebius	9.50 dollars per GPU-hour, i.e. 76 dollars per hour for 8 GPUs	nebius.com/prices
Purchase of a Supermicro SYS-822GS-NB3RT server (8 × HGX B300)	from 545,173.05 dollars, starting configuration	Supermicro store
Power draw of an 8 × B300 system	about 14 kW (NVIDIA DGX B300)	NVIDIA DGX B300 specifications
Electricity, French businesses using 20 to 499 MWh per year	0.2049 euros per kWh, excluding VAT and recoverable taxes (first half of 2026)	Eurostat, nrg_pc_205
Exchange rate	1 euro = 1.1193 dollars (reference rate of 23 October 2026)	European Central Bank
K3 throughput on one 8 × B300 node	1,395 tokens per second per GPU, input and output counted together	SGLang Kimi-K3 cookbook, published measurements

Throughput is the most fragile input. The SGLang team measured it on one 8 × B300 node (a configuration marked "Verified"), with synthetic requests of 8,192 input tokens and 1,024 output tokens, and 64 concurrent requests. In that regime the first token arrives after 11.6 seconds, then one token every 40 milliseconds. That is batch throughput, not an interactive assistant. On agent traffic with a latency target, the author of the 8 × B300 deployment cited above measures a sustainable rate of 0.5 to 0.7 requests per second. For the MI355X, the cookbook does not publish a verified measurement yet, so the scenarios below cover the B300 only.

Assumptions, to replace with your own

• Request shape: 8,192 input tokens for 1,024 output tokens, a ratio of 8 to 1. That is the benchmark shape, close to RAG or agent use (long context, short answer). If your answers are longer, the API cost rises and the node's throughput changes.
• Length of a month: 730 hours, node running continuously.
• Load: the node runs at benchmark throughput. Its capacity is then about 29.3 billion tokens per month (11,160 tokens per second × 3,600 × 730). That is an optimistic ceiling: real load follows your working hours and your peaks.
• Purchase: straight-line depreciation over 3 years (36 months), with no financing and no resale value.
• Electricity: 14 kW continuously, without the extra cost of cooling.
• Left out: data-centre hosting, network, storage, operations staff, support beyond the warranty, software, shipping and import costs, VAT. The purchase cost in the table is therefore a floor.
• No API caching: K3.1 Pro charges 0.20 dollars per million cached input tokens. If your prompts reuse a long shared prefix, the API costs less than in this calculation.

The formulas, to redo the calculation with your prices

monthly API cost       = (input tokens × input price + output tokens × output price) / 1,000,000
monthly rental cost    = node hourly price × billed hours
monthly purchase cost  = server price / months of depreciation
                         + power (kW) × hours × price per kWh
                         + hosting and operations
node capacity          = throughput (tokens/s) × 3,600 × hours of use
break-even volume      = monthly node cost / average API price per million tokens

Under our assumptions, the average price of one million tokens through K3.1 Pro is (8,192 × 2 + 1,024 × 10) / 9,216 = 2.89 dollars, or 2.58 euros. IgnitionAI estimate.

The scenarios

Monthly cost excluding taxes, in euros, at 24 October 2026 prices. IgnitionAI estimate.

Monthly volume	Equivalent requests per day	K3.1 Pro API	Massed Compute rental	Scaleway rental (Paris)	Nebius rental	Purchase over 3 years + electricity (floor)
100 million tokens	about 360	€260	€34,400	€43,800	€49,600	€15,600
1 billion	about 3,600	€2,600	€34,400	€43,800	€49,600	€15,600
5 billion	about 17,800	€12,900	€34,400	€43,800	€49,600	€15,600
10 billion	about 35,700	€25,800	€34,400	€43,800	€49,600	€15,600
20 billion	about 71,300	€51,600	€34,400	€43,800	€49,600	€15,600
29 billion (full node)	about 103,400	€74,800	€34,400	€43,800	€49,600	€15,600

IgnitionAI estimate. One request is 9,216 tokens (8,192 + 1,024). Rental: hourly price × 730 hours, converted at the ECB rate of 23 October 2026. Purchase: 545,173.05 dollars, or €487,066, divided by 36 months (€13,530), plus 14 kW × 730 hours × €0.2049 (€2,094). Beyond about 29.3 billion tokens per month, you need a second node.

Break-even points against the K3.1 Pro API, again as an IgnitionAI estimate:

• Purchase: from about 6 billion tokens per month, 21% of the node's capacity. That is a floor, before hosting and staff.
• Massed Compute rental: about 13.3 billion tokens per month, 45% of capacity.
• Scaleway rental in Paris: about 17.0 billion, 58% of capacity.
• Nebius rental: about 19.2 billion, 65% of capacity.

What these numbers say

Below a few billion tokens per month, the API wins by a wide margin. One billion tokens per month already means 3,600 requests of 9,000 tokens every single day. According to an IgnitionAI estimate, most mid-market companies use far less.

Rental only beats the API if the node runs at least half full, day and night, with traffic close to the benchmark. If your use is interactive and concentrated in office hours, the node spends much of the month waiting, and the break-even moves further out.

Purchase becomes competitive earlier on paper, around 6 billion tokens per month. But that figure leaves out hosting, cooling and the team that keeps the service running. Supermicro also reports longer than usual lead times.

The API price matters as much as the hardware price. With K3 at the price displayed by OpenRouter (0.239 dollars input, 10.80 dollars output), the average price falls to €1.26 per million tokens. The purchase break-even then moves to 12.4 billion tokens per month, Massed Compute to 27.3 billion, and the Scaleway and Nebius rentals exceed the capacity of one node. Conversely, against K3 served by Moonshot AI (3 and 15 dollars, or €3.87 per million), the break-even points drop to 4.0 billion for purchase and 8.9 billion for Massed Compute. IgnitionAI estimate.

One last point: these scenarios compare K3, whose weights are published, with the API of K3.1 Pro, which ranks higher (55 points on the Artificial Analysis Intelligence Index, against 44 for K3). Hosting K3 today therefore means paying more for a lower-ranked model than the one behind the API, until the K3.1 weights are published.

The reasons that are not financial

Dedicated hosting can be justified without being cheaper. Your data stays on infrastructure you choose, in France with a node rented from Scaleway in Paris, or on your premises with a server you buy. The model stays frozen in the version you validated. These arguments matter for regulated sectors, but they have a price, which the table above helps you put a number on. And where the computation runs does not settle GDPR questions on its own, see below.

What it actually changes for a mid-market company

1. Reversibility becomes a technical fact rather than a contractual promise. A proprietary model can change price, version or terms overnight. A model whose weights are published stays available in the version you validated. You can host it yourself, have a third party host it, or go back to the API. That optionality is a negotiating argument with your current suppliers, and a concrete answer to a risk committee questioning your dependency.

2. Sovereign hosting becomes attainable. Your data can be processed by a high-capability model hosted on European infrastructure, with no transfer to a vendor outside the European Union. For regulated sectors, that shifts a common constraint, according to an IgnitionAI estimate. The topic still needs work with your DPO: where the compute sits does not by itself settle questions of legal basis and processing under the GDPR.

3. The market's price floor drops. The existence of a capable open model puts pressure on every provider's pricing. Your inference budgets, which we break down in our article on the cost of a RAG project, benefit from that competition whichever model you end up choosing.

So the right question is not "open or proprietary". It is "which model for which task". A production system routes simple requests to an economical model and complex ones to a more capable model, with continuous quality evaluation. The choice is made task by task, on your data, not on a general leaderboard.

Decision grid

Your situation	Recommended approach
Moderate volume, no localisation constraint	API, comparing providers. Open weights give you negotiating leverage, not an infrastructure project
Sensitive data, sovereignty constraint	Host the open model with a European provider, or on your own infrastructure if your teams can operate it
Industrial, continuous volume	Compute the break-even with your real volumes. This is the only case where self-hosting defends itself economically
Product resold to third parties	Read the licence before any commitment, in particular the commercial thresholds and the attribution requirement

IgnitionAI estimate. Possible variation depending on context.

FAQ: Kimi K3 and open models in the enterprise

• Is Kimi K3 open source?

  Not strictly. Its weights are published for free download, but the Kimi K3 License is not an open-source licence. It requires a separate agreement with Moonshot AI for model-as-a-service businesses exceeding 20 million US dollars over twelve months. It also requires displaying the "Kimi K3" attribution for products exceeding 100 million monthly active users or 20 million US dollars in monthly revenue. Internal use not made available to third parties is exempt.

• Can you host Kimi K3 on your own servers?

  Technically yes, but not on ordinary hardware. At the native four-bit quantisation, the 2.8 trillion parameters occupy around 1.4 TB of GPU memory on paper, and the published checkpoint weighs 1.56 TB, before the attention cache. That means a node of 8 GPUs with 288 GB each, such as B300 or MI355X. IgnitionAI estimate: dedicated hosting is justified by sovereignty or industrial volume, rarely by economics alone against an API price of around 2.90 dollars per million input tokens for K3 (August 2026), and 2 dollars for K3.1 Pro (October 2026).

• Is it cheaper to host Kimi K3 on your own servers than to use the API?

  Rarely, at 24 October 2026 prices. IgnitionAI estimate, with requests of 8,192 input tokens and 1,024 output tokens: a rented 8 × B300 node only becomes cheaper than the K3.1 Pro API (2 and 10 dollars per million tokens) beyond 13 to 19 billion tokens per month, depending on the provider. A purchased node depreciated over 3 years drops below the API around 6 billion tokens per month, before hosting and an operations team. Hardware prices move fast: redo the calculation with your own quotes, using the formulas in the article.

• Are open models as good as proprietary ones?

  On published evaluations, Kimi K3 sits at a high level, and Moonshot itself states that its overall performance still trails the most powerful proprietary models. Its successor, Kimi K3.1 Pro, launched on 23 October 2026, narrows the gap: Artificial Analysis ranks it as the 4th model on its Intelligence Index, with 55 points. Its weights are announced for 4 November 2026 at the latest but are not published yet, and its licence is not known yet. The useful decision criterion is not a general leaderboard but an evaluation on your own data, task by task.

• What does publishing the weights change for a mid-market company?

  Three things. Reversibility becomes a technical fact, since the model stays available in the version you validated, hostable by you or by a third party. Sovereign hosting on European infrastructure becomes attainable for regulated sectors. And competitive pressure on pricing benefits your inference budgets, whichever model you choose.

Methodology and sources

Primary sources (accessed 2 August 2026):

• Official model card, specifications and evaluations: huggingface.co/moonshotai/Kimi-K3
• Moonshot AI technical blog, architecture and positioning: kimi.com/blog/kimi-k3
• Original LICENSE file, commercial thresholds and exemptions: huggingface.co/moonshotai/Kimi-K3, LICENSE
• Reference API pricing: openrouter.ai/moonshotai/kimi-k3

Update of 24 October 2026 (sources accessed on 24 October 2026):

• Moonshot AI blog, Kimi K3.1 Pro launch and announced weights date: kimi.ai/blog/kimi-k3-1-pro
• Artificial Analysis, Intelligence Index, ranking, weights status and K3.1 Pro pricing: launch article and model leaderboard
• Size of the published K3 checkpoint: Hugging Face API, moonshotai/Kimi-K3
• Serving K3 on one node of 8 MI355X: vLLM recipe
• K3 deployment on 8 B300, B300 and H200 memory: vincentzed.com/writing/kimi-k3-deployment
• Memory of an 8 × B200 server: NVIDIA DGX B200 1440GB listing, Delta
• Memory of an AMD MI355X GPU (288 GB): AMD Instinct MI355X product page
• Memory of an 8 × MI300X server: Supermicro MI300X listing, Gatewell
• K3.1 Pro API pricing: openrouter.ai/moonshotai/kimi-k3.1-pro

Hosting cost scenarios (sources accessed on 24 October 2026):

• K3 throughput on 8 B300, verified measurements: SGLang Kimi-K3 cookbook and measurements file
• K3 provider pricing: OpenRouter API, moonshotai/kimi-k3
• 8 × B300 rental: Massed Compute, Scaleway public catalogue and Scaleway GPU pricing, Nebius
• Purchase price of an 8 × HGX B300 server: Supermicro store
• Power draw of an 8 × B300 system: NVIDIA DGX B300 specifications
• Electricity prices for businesses in France: Eurostat, nrg_pc_205
• Euro-dollar exchange rate: ECB reference rates

Regulatory: Regulation (EU) 2016/679 (GDPR), for processing and transfer questions: EUR-Lex.

IgnitionAI calculations and estimates: the memory footprint of roughly 1.4 TB follows from the direct calculation of 2.8 trillion parameters at four bits, excluding the attention cache. The published checkpoint weighs 1.56 TB according to the Hugging Face API (sum of the files, accessed on 24 October 2026). Infrastructure cost orders of magnitude and the per-situation recommendations are an IgnitionAI estimate. The monthly costs, depreciation, node capacity and break-even points in the "Rent, buy or stay on the API" section are IgnitionAI calculations based on prices collected on 24 October 2026 and on the assumptions listed in that section. Possible variation depending on context. The evaluation scores cited are self-reported by the vendor and have not been reproduced by IgnitionAI.

Independence: IgnitionAI has no commercial relationship with Moonshot AI, nor with the other model vendors cited. See our editorial policy.

Sources last reviewed: 24 October 2026. API pricing and model licences change quickly, re-check at scoping time.

Related IgnitionAI articles:

• How much does an enterprise RAG project cost in 2026
• Why AI projects don't reach production
• Enterprise RAG in production: 5 critical decisions your first POCs hide from you

---

Wondering which model fits your case, and whether sovereign hosting is justified? Our two-week scoping compares the options on your real data, prices the inference costs and ends with a written go/no-go. Request a conversation.