Kimi K3: what a frontier-class open model changes for your sovereignty and your bill
Moonshot AI released the Kimi K3 weights on 27 July 2026: 2.8 trillion parameters, 104 billion active, a 1-million-token context window. For most mid-market companies this changes nothing about self-hosting: the weights alone take around 1.4 TB of GPU memory. It changes three other things: reversibility, the market's price floor, and sovereign hosting. An analysis of the licence, the break-even maths and the decision grid. For CIOs, CTOs and innovation leaders.
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. 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 occupy around 1.4 TB of 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.
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. That is the memory for the weights alone, before the attention cache, which grows with context length and concurrent users.
No single GPU server provides that memory today. You need a datacentre-class multi-GPU node, and more still to exploit the full one-million-token window. 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 is priced at around 2.90 dollars per million input tokens and 14 dollars per million output tokens (OpenRouter pricing, accessed 2 August 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.
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 constraint we meet on nearly every engagement. 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 estimates based on our 3 engagements in 2024-2025 for French mid-market companies in regulated sectors. 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, before the attention cache. That means a datacentre-class multi-GPU node. 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.
-
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. The gap has narrowed markedly. 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
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. Infrastructure cost orders of magnitude and the per-situation recommendations rest on 3 engagements delivered in 2024-2025 for French mid-market companies in regulated sectors. 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: 2 August 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.