
Moonshot AI says the full Kimi K3.1 Pro weights ship by 4 November 2026. For a European buyer, those open weights buy rights, not a model in the hospital basement: a 3T-class model needs roughly ten H200 GPUs just to hold its weights. What “open” gives you is the right to inspect, fine-tune, distill and self-host on a cluster you control – under licence terms that have become policy.
I have written the “open crosses the frontier” story before, for GLM-5.2’s open-weight frontier claim and for Kimi K3. This is the buyer’s reading instead: what openness is worth to a hospital, bank or public agency in the EEA that will never run a 2.8-trillion-parameter model on one server. My thesis: in 2026 “open” is a licence design, not a property of the weights, and procurement should treat it that way.
- Release date: “Full model weights will be released by Nov 4, 2026,” per Moonshot’s K3.1 Pro announcement (23 October 2026).
- Scale: “3T-class” per Moonshot; Artificial Analysis lists 2.8 trillion parameters and a 1-million-token context window.
- Price: $2 per million input tokens and $10 per million output tokens on the API – half the price of Claude Opus 5.5.
- Benchmark: 65.7 % on Terminal-Bench 4.0 against 66.4 % for Claude Opus 5.5 (vendor-reported).
- Licence: not named for K3.1 as of 25 October 2026; Kimi K3’s licence requires a separate agreement for model-as-a-service businesses above $20 million in revenue over 12 months.
- European counterpoint: Aleph Alpha’s Kolibri-1 (78B, Apache 2.0) fits on a single NVIDIA H200.
When will the Kimi K3.1 Pro weights be released – and what exactly is promised?
Moonshot’s wording is precise: the model is available now on Kimi.ai, Kimi Work, Kimi Code and the Kimi API, and “full model weights will be released by Nov 4.” The company describes Kimi K3.1 Pro as “a post-trained version of Kimi K3, with substantially larger post-training compute” – the same base, with a lot more work done on top.
Three things are missing. The announcement names no licence, gives no active-parameter count (Kimi K3 uses 104 billion active parameters per token, per its Hugging Face model card) and discloses no training compute. As of 25 October, Artificial Analysis still lists the licence as “Proprietary” with the weights not public.
Dates like this slip. Mistral announced Mistral Large 4 on 6 October 2026 and promised “the weights by the end of the month”; on 25 October, neither its news page nor its Hugging Face organisation had them. Treat “by 4 November” as a commitment to write into a contract, not a delivered asset.
How close is Kimi K3.1 Pro to Claude Opus 5.5?
Close enough that the gap no longer decides anything. On Terminal-Bench 4.0, Moonshot reports 65.7 % for K3.1 Pro, 66.4 % for Claude Opus 5.5 and 58.2 % for GPT-6 Astra, run at maximum reasoning effort. On BenchCAD with tools, K3.1 Pro trails both (92.7 % against 96.2 % for Opus 5.5). All vendor-reported; a 0.7-point gap on one benchmark is noise.
The independent reading is more useful. Artificial Analysis scores K3.1 Pro at 55 on its Intelligence Index, fourth of 215 models, three points behind Claude Opus 5.5 at 58. Simon Willison (23 October 2026) called it “the first time a (soon-to-be) open weight model has scored this highly against the frontier”. He also noted the price: $2/$10 per million tokens, “half the price of Claude Opus 5.5”.
One caveat: Artificial Analysis counted 150 million output tokens for K3.1 Pro on its index runs, against a median of 87 million. Verbosity eats part of the price advantage, so measure cost per completed task.
A detail for your security officer: Moonshot says it “identified and removed cyber-related data from both training stages”. On its xdev-bench-hard test, the released model scores 81.0 against 90.5 for a checkpoint taken before that mitigation. The downloadable weights are the mitigated version.
Can a hospital run a 3-trillion-parameter model on-premises?
Not in any server room I have seen. By my arithmetic, 2.8 trillion parameters at 4 bits – the precision Kimi K3 ships in – come to roughly 1.4 TB for the weights alone. NVIDIA specifies the H200 with 141 GB of HBM3e memory, so you need about ten of them before you have stored a single token of context. That is more than one eight-GPU node, before the KV cache for a 1-million-token window. Mixture-of-experts saves compute, not memory. I worked through this 1.4-terabyte question when Kimi K3 launched, and K3.1 does not change it.
In clinical AI, the first question a hospital asks me is never the benchmark. It is where the weights run and who can change them. For a 3T-class model, the honest answers are:
- A dedicated cluster in an EEA data centre, single-tenant, under your own contract. Real sovereignty, but a capex and operations commitment, not a download.
- Distillation or fine-tuning of a smaller model, using the big one as teacher on your own infrastructure. This is where the open weights do the most work.
- A smaller open model that fits the hardware you already have.
Europe has an answer for the third. Aleph Alpha’s Kolibri-1, released on 3 October 2026, is a 78-billion-parameter mixture-of-experts model with 3.46 billion active parameters, licensed under Apache 2.0. Its FP8 weights take about 78 GB, and Aleph Alpha lists a single H200 as a minimum configuration. It is not a frontier model; its published comparisons with Qwen are mixed. But it is a model a regional health trust can actually run behind its own firewall.
Why has “open” become licence policy?
Read the licences rather than the press releases. The Kimi K3 License starts as an MIT-style grant, then adds conditions. A licensee running a model-as-a-service business whose aggregate revenue with its affiliates exceeds $20 million “over any consecutive 12 months” must sign a separate agreement with Moonshot. Products with more than 100 million monthly active users, or more than $20 million in monthly revenue, must display “Kimi K3” prominently.
Meta’s Muse Model License for Muse Spark Open, dated 15 October 2026, uses the same design. A model-as-a-service provider with more than $20 million in gross revenue over any 12-month period must notify Meta and sign a separate written agreement. It also requires a “Built with Muse Spark Open” notice and attaches a usage policy. For EU licensees, the counterparty is Meta Platforms Ireland Limited.
Two competitors, one clause: free for internal use and research, a contract for anyone who resells the model as an API. That is a sensible business model, but not open source in the OSI sense. By contrast, DeepSeek-V4.1-Flash ships under a plain MIT licence.
For a hospital, the revenue clauses will rarely bite. The definition of “internal use” matters more. The K3 licence defines it as any use that does not make the software, “its outputs, or its underlying capabilities available to third parties”. A patient portal built on the model is not obviously internal. That is the wording your legal team should read before an architect builds anything.
The closed side has turned licence policy into enforcement. On 12 October 2026, Anthropic published its distillation defences. It says its new methods have “reduced the amount of data harvested by distillers by around 90 to 99%”. It also says it found a case in which “requests to Moonshot’s Kimi models were secretly forwarded to Claude” as part of a distillation effort. Those are one competitor’s allegations; I found no response from Moonshot. The structural point stands either way: closed vendors police distillation, open-weight licences grant it under conditions. For a buyer, the legal right to distill is among the most valuable things “open” still means.
Does an open-weight model get a pass under the EU AI Act?
Partly, and probably not this one. Article 53(2) waives two duties for general-purpose models: technical documentation, and information for downstream providers. It covers models under “a free and open-source licence that allows for the access, usage, modification, and distribution of the model”, with public weights. The copyright policy and training-data summary still apply, and the article adds: “This exception shall not apply to general-purpose AI models with systemic risks.”
Under Article 51(2), a model is presumed to carry systemic risk when its cumulative training compute exceeds 10^25 FLOPs. Moonshot has not disclosed K3.1 Pro’s compute, but a 3T-class frontier model is, in my reading, almost certainly in that territory. Whether a licence with revenue-triggered contracts counts as “free and open-source” is a second question the AI Office has not settled. I covered what already applies in my piece on the AI Act omnibus and October 2026, including the fact that Norway has not yet incorporated the Act.
The practical consequence: “open weights” never replaces documentation in your supplier file.
What should a European buyer require when a vendor calls a model “open”?
This is the checklist I now use, on top of the selection criteria in how I choose LLMs for production:
- The licence text itself, not a summary: OSI-approved (Apache 2.0, MIT) or custom?
- Commercial triggers: revenue, MAU or model-as-a-service thresholds, and how “internal use” is defined.
- Attribution and usage-policy obligations, including whether the vendor can change the usage policy later.
- Counterparty and jurisdiction: which legal entity you contract with and which law governs the contract.
- AI Act artefacts: the training-data summary, the copyright policy, and a statement on systemic-risk status and training compute.
- Distillation and fine-tuning rights, stated explicitly, including rights to derivative models.
- Version pinning: hashes of the weights you evaluated, so the model in production is the model you validated.
- A hardware bill of materials for your context length and throughput, not the vendor’s demo.
The same holds for closed products – see why always-on agents stop at the EEA border. Either way, the paperwork is the product.
What I would take from this
Kimi K3.1 Pro shows that the capability gap between open and closed is now small. The gap that matters to regulated buyers lies in licence terms, documentation and hardware. My predictions:
- By 4 November 2026, the K3.1 Pro weights will be on Hugging Face under a licence with a model-as-a-service revenue clause much like K3’s. They will not ship under plain MIT.
- By 30 June 2027, at least one large European health or public-sector buyer will publicly choose a distilled or sub-100B open model over a 3T-class one, with on-premises hosting as the stated reason.
- Before the end of 2027, the AI Office or a national authority will publish guidance on whether custom licences with commercial triggers qualify for the Article 53(2) exemption.
Frequently asked questions
When will the Kimi K3.1 Pro weights be released?
Moonshot AI says the full weights will be released by 4 November 2026. As of 25 October 2026, they were not yet public, and the licence had not been announced.
Can a 3-trillion-parameter open-weight model be run on-premises in a hospital?
Not on a single server. At 4-bit precision, roughly 1.4 TB of weights need about ten 141 GB H200 GPUs before any context is stored. Realistic paths are a dedicated cluster in an EEA data centre, distillation into a smaller model, or a model such as Kolibri-1 that fits on one GPU.
Is Kimi K3.1 Pro better than Claude Opus 5.5?
Not overall. It is 0.7 points behind on Moonshot’s Terminal-Bench 4.0 figures, three points behind on the Artificial Analysis Intelligence Index (55 vs 58), and half the price per token.
What licence will Kimi K3.1 Pro use?
Moonshot has not said. Its predecessor used the Kimi K3 License, an MIT-style grant with a $20 million model-as-a-service clause and attribution requirements.
What should a European buyer require when a vendor calls a model “open”?
The full licence text with its commercial triggers and definition of internal use, the counterparty and jurisdiction, and the AI Act documentation – plus explicit distillation rights, pinned weight hashes and a hardware bill of materials for your workload.