🔍 Read the full analysis: Open-Weight AI Is Near, But On-Premises Needs Volume on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
IBM CEO Arvind Krishna says open-weight models can deliver about 90% of frontier-model performance at a fraction of the token cost, making on-premises AI attractive to some enterprises. Artificial Analysis data read Oct. 26 put the leading downloadable model at 80% of the proprietary leader; the economics of buying hardware depend heavily on usage, staffing and other costs.
IBM CEO Arvind Krishna said on the company’s October 21 earnings call that open-weight models can cost 10% to 25% as much per token as frontier models while being “90% as effective.” But the latest Artificial Analysis comparison cited here puts the best downloadable model at 80% of the proprietary leader, and the cost case for running AI on a company’s own servers depends on keeping expensive hardware busy.
Krishna said enterprises could begin on-premises deployments with two or three racks. He argued that many routine tasks—such as summarizing email, examining contracts and answering customer calls—do not require the most capable available model. He also said banks and trading firms may want control over data and logs, and that open weights provide rights to run models on private infrastructure, subject to each model’s licence.
The comparison from Artificial Analysis, read October 26, gives Claude Opus 5.5 a score of 57.6 and Xiaomi’s downloadable MiMo-V2.6-Pro a score of 46.3. That makes the open model about 80% of the proprietary leader on the index, which uses ten evaluations. The article’s historical figures show the open-to-closed share at 51% at the end of 2024 and 74% at the end of 2025. Those comparisons indicate gains, but they are not the same as confirming Krishna’s 90% figure on a named benchmark.
Hardware costs also complicate the pitch. The source lists an eight-GPU H200 server from $172,643 and AWS rental of the same GPU class at $63.30 an hour. Its estimate says purchase costs would match rental in about 114 days at full utilization, or around 15 months at 25% utilization, before staffing, cooling and networking. Those are illustrative list-price calculations, not a complete cost comparison.
IBM's 90% claim, checked
Open weights reach 80% of the leader
Krishna says open-weight models are "90% as effective". On an independent index the best model you can download today reaches 80% of the best closed one, up from about half two years ago.
Best open model as a share of the best closed
Artificial Analysis Intelligence Index, current version, best downloadable vs. best proprietary model at each date.
What Krishna said
Illustrations from the earnings call, not a named benchmark.
- Late 2024"70" vs. "20 or 30"Krishna, Q3 call, 21 Oct · no named benchmark
- Late 2025"70 to 50"Krishna, Q3 call, 21 Oct · no named benchmark
- Now"90% as effective"Krishna, Q3 call, 21 Oct · no named benchmark
Top downloadable models today
Intelligence Index score and licence type. Read the licence before the leaderboard.
| Model | Index | Licence |
|---|---|---|
| MiMo-V2.6-Pro Xiaomi | 46.3 | MIT |
| DeepSeek V4.1 Flash 1017 DeepSeek | 45.1 | MIT |
| GLM-5.3 Z AI | 44.8 | own commercial licence |
| Step 5 Preview StepFun | 43.7 | own commercial licence |
| Muse Spark Open Meta | 38.9 | licence with conditions |
In points the gap did not shrink: 7.5 points at the end of 2024, 11.3 today. Closed models moved too.
Sources: Artificial Analysis, Intelligence Index v4.3.2 (read 26 Oct 2026), IBM Q3 2026 earnings call transcript (21 Oct 2026), Artificial Analysis on Kimi K3.1 Pro (23 Oct 2026).
Powered by Thorsten Meyer AI
The Economics Depend on Utilization
The debate is not simply whether open models are cheaper per token. For companies considering their own infrastructure, the decision also involves upfront hardware spending, electricity, cooling, networking and staff who can operate and optimize the systems. A server that runs near capacity may compare favorably with renting, while idle capacity stretches the payback period.
Krishna acknowledged a skills constraint, asking where enterprises would find kernel and network engineers. He said that gap remains except at the most sophisticated organizations, while cloud services are the easier option despite being more expensive. That caveat matters: model weights do not provide operating expertise, and a low token price does not by itself establish lower total cost.
For organizations handling sensitive information, the appeal also includes greater control over where data and logs reside. But on-premises deployment is not the only route to private use: Krishna also cited private clouds. Buyers must weigh the relevant model’s performance and licence alongside infrastructure and governance requirements.
Products Worth Considering
Jackery Explorer 2000 v2 Portable Power Station with Two 200W Solar Panels
Jackery Explorer 300 Portable Power Station, 292Wh
Logitech MX Keys S Wireless Keyboard, Graphite
As an affiliate, we earn on qualifying purchases.
Benchmarks, Licences and Hardware
Krishna’s “90%” comparison was offered in an earnings-call answer as a broad illustration, rather than tied to a specific benchmark. He described a narrowing capability gap, saying that in late 2024 an open model might score 20 or 30 against a leading model’s 70, and that by late 2025 the comparison was closer to 50 against 70. Artificial Analysis provides one independent comparison, but its index score is not a universal measure of performance across every task.
The source says that Kimi K3.1 Pro scored 55.1, close to Claude Opus 5.5’s 57.6. However, Moonshot had only promised to release the weights “by Nov. 4.” Until those weights are available, the model cannot be counted as downloadable open-weight software. The source also says no licence had been published, another reason not to treat the score as proof of practical availability.
Licensing varies across models. The source identifies MiMo-V2.6-Pro and DeepSeek V4.1 Flash 1017 as MIT-licensed, while other models use custom commercial or community terms. A high leaderboard position does not settle whether a model can be used for a particular business purpose; organizations need to check the licence and deployment conditions.
The 90% Figure Has Limits
The available figures do not establish that open-weight models are consistently 90% as capable as proprietary frontier models. The Artificial Analysis index provides a comparison across ten evaluations, but performance can differ by task, and Krishna did not name the benchmark behind his estimate. The comparison also changes as closed models improve.
Kimi K3.1 Pro’s weights were not yet available in the source material, and its licence was unpublished. Its score therefore describes a model whose practical open-weight status remained conditional. The hardware payback figures likewise omit operating expenses and assume particular utilization rates; the material does not provide a full, independently verified total-cost analysis or establish how representative the listed prices are.
IBM has a commercial interest in on-premises infrastructure and sells hardware, including Spyre cards. Its earnings-call claims should be read with that context in mind. The source also says IBM described its llm-d project as reducing costs by five to ten times, but provides no independent measurement to verify that claim.
Watch for Weights and Real-World Costs
The next concrete test is whether Moonshot releases Kimi K3.1 Pro’s weights as promised by November 4, and under what licence. That would clarify whether its near-leading score applies to a model enterprises can actually download and deploy.
For buyers, the practical next step is to compare models on their own workloads and calculate infrastructure costs using realistic utilization, staffing and energy assumptions. Further benchmark updates may show whether open models continue to close the capability gap, while real deployment data would be needed to test the advertised savings. Until then, lower token prices are only one part of the on-premises case.
Key Questions
What did IBM’s CEO claim about open-weight models?
Arvind Krishna said open-weight models can cost 10% to 25% as much per token as frontier models while being “90% as effective.” He presented this as a broad comparison on IBM’s October 21 earnings call, not as the result of a named benchmark.
How close is the leading downloadable model to the top proprietary model?
In the Artificial Analysis figures cited as read October 26, 2026, MiMo-V2.6-Pro scored 46.3 against Claude Opus 5.5’s 57.6, or about 80% of its score on that index.
Why does on-premises AI need high usage?
Companies pay for hardware whether or not it is busy. The cited estimate puts the purchase-versus-rental crossover at about 114 days at full utilization, compared with roughly 15 months at 25% utilization, before staff and other operating costs.
Is Kimi K3.1 Pro available as an open-weight model?
Not according to the source material dated October 26, 2026. Moonshot had promised weights by November 4, but they had not shipped, and the source said no licence had been published.
Does a high benchmark score mean a model is suitable for business use?
No. Organizations also need to test the model on their own tasks and review its licence, deployment requirements, privacy implications and operating costs. Index scores alone do not settle those questions.
Source: ThorstenMeyerAI.com
