🔍 Read the full analysis: IBM’s CEO Sees Open-Weight Models Approaching On-Prem Use 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 said open-weight models can cost 10% to 25% as much as frontier models while being about 90% as effective, making on-premise deployment worth considering for some enterprises. An Artificial Analysis comparison read Oct. 26 put the leading downloadable model at 80% of the top proprietary model; a higher-scoring model has not yet released its weights.
IBM CEO Arvind Krishna said open-weight AI models can cost 10% to 25% as much as frontier models while being “90% as effective,” making it practical for some large companies to run AI on their own infrastructure. But a model comparison available on Oct. 26 put the best downloadable model at 80% of the leading proprietary model, leaving a larger performance gap than Krishna’s estimate suggests.
Krishna made the comments during the question-and-answer portion of IBM’s Q3 earnings call on Oct. 21. He described open models as a lower-cost option for routine business work, including summarizing email, examining contracts, answering customer calls and composing messages. A large enterprise, he said, could begin with two or three racks of equipment. He also argued that open weights give customers permission and intellectual-property rights to run models on premises or in private clouds.
The cost and capability figures are Krishna’s characterization, not results from a benchmark he named on the call. Artificial Analysis, which compares models using a set of evaluations, listed Xiaomi’s downloadable MiMo-V2.6-Pro at 46.3 and proprietary Claude Opus 5.5 at 57.6 on its index as of Oct. 26. That puts MiMo at about 80% of the leading score. The source’s historical comparison puts the best open model at 51% of the proprietary leader at the end of 2024 and 74% at the end of 2025.
A possible exception is Kimi K3.1 Pro, which scored 55.1 on the index, or about 96% of Claude Opus 5.5’s score. But its developer, Moonshot, had promised the weights by Nov. 4; they had not shipped by the date of the comparison. Until weights are available, the model is not an open-weight option for companies to download and run. The source also notes that no licence had been published and that the model’s reported 2.8-trillion-parameter scale would make deployment a substantial infrastructure undertaking.
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 Cost of Running Models On-Site
For companies handling sensitive records, the case for on-premise AI is about control as well as price. Krishna cited banks and traders that want data and usage logs to remain within their control. Downloadable weights can give customers more choice over where inference runs, though they still need suitable hardware, technical staff and a licence that permits the intended use.
The economics depend on how often the hardware is used. The source compares an eight-GPU H200 server listed from $172,643 with renting the same GPU class on AWS for $63.30 an hour. Its arithmetic estimates that buying pays off in about 114 days at full utilization, but about 15 months at 25% utilization, before accounting for staffing, cooling or networking. These are estimates based on listed prices, not a universal cost comparison; workload, discounts and operating expenses can change the result.
Krishna acknowledged that technical capacity can be a constraint, asking where enterprises would find kernel engineers and networking specialists. He said that gap remains outside the most sophisticated companies. That caveat matters: lower token prices do not by themselves make on-prem deployment cheaper if a company cannot keep its systems well utilized or lacks the people to operate them.
Popular right now
TCL 85 Designer Series QD-Mini LED TV (Model 85A400U) with Google Smart TV
Bose QuietComfort Headphones (2nd Gen), Wireless Noise-Cancelling Over-Ear Headphones, Eucalyptus Green
Breville Bambino Espresso Machine, Brushed Stainless Steel
As an affiliate, we earn on qualifying purchases.
IBM’s Pitch and the Model Gap
Krishna said the performance gap had narrowed over time, describing a broad comparison in which open models moved from scores around 20 or 30 against a proprietary model at 70 in late 2024, to roughly 50 against 70 in late 2025. Those figures were presented as illustrative examples, not as results from a specified evaluation. The Artificial Analysis figures in the source support the direction of improvement, while showing that the leading open model’s score is still below 90% of the proprietary leader.
The source also points to an important distinction between open weights and open-source licensing. Models can make weights downloadable but still attach conditions to commercial use. MiMo-V2.6-Pro and DeepSeek V4.1 Flash 1017 are identified as MIT-licensed in the source; other models use separate commercial or custom licences. Businesses need to review the applicable terms rather than infer usage rights from a model’s presence on a leaderboard.
IBM has a commercial interest in a shift toward on-premise AI infrastructure. Krishna said demand could support sales of the company’s Spyre cards, and IBM also claimed in prepared remarks that its llm-d project can run large models at five to 10 times lower cost. That figure is IBM’s claim; the source says it could not independently verify it. The article also reports that IBM’s mainframe revenue fell 20% in the quarter, based on The Register’s account, adding to the business context around the company’s infrastructure pitch.
Costs, Licences and Weights
The available comparison does not settle whether open-weight models are generally “90% as effective.” Artificial Analysis scores cover a defined set of evaluations; they do not establish that models perform equally across every enterprise task, language or deployment setting. Krishna did not identify a benchmark behind his percentage, and the source does not provide enough detail to validate the token-cost estimate across providers and workloads.
It also remains unclear whether Moonshot will release Kimi K3.1 Pro’s weights by the promised date, what licence would apply, and what hardware and operating costs a deployment would require. The server-versus-rental figures exclude staff, cooling and networking, while actual prices can differ from list prices. IBM’s stated cost reduction for llm-d has not been independently measured in the source.
Watch for Kimi’s Weight Release
The near-term test is whether Moonshot publishes Kimi K3.1 Pro’s weights by its stated Nov. 4 target, and under what licence. If it does, the model’s score would narrow the measured gap against Claude Opus 5.5, but companies would still need to determine whether the licence and hardware requirements fit their needs.
For enterprises weighing on-premise AI, the next step is likely to be workload-specific comparison: test model quality on actual tasks, verify commercial rights, and calculate total infrastructure and staffing costs at realistic utilization. IBM’s argument is a forecast about adoption, not confirmation that companies are already moving workloads on-site at scale. Further deployment data and independently checkable cost comparisons would help show how far the economics extend beyond the examples Krishna gave.
Key Questions
What did IBM’s CEO say about open-weight models?
Arvind Krishna said open-weight models can cost 10% to 25% as much as frontier models while being about 90% as effective. The percentage is his estimate, not a result tied to a benchmark he named on the call.
How close are the leading downloadable models to proprietary ones?
In the Artificial Analysis comparison read Oct. 26, 2026, MiMo-V2.6-Pro scored 46.3, compared with 57.6 for Claude Opus 5.5, or about 80% of the proprietary model’s score. The index measures performance on its own evaluation set.
Is Kimi K3.1 Pro available to run on company hardware?
Not according to the source’s Oct. 26 status. Moonshot had promised the weights by Nov. 4, but they had not yet been released. Its licence and deployment requirements were also unresolved.
Does on-premise AI automatically cost less?
No. The source’s example estimates that buying an eight-GPU H200 server could pay off in about 114 days at full utilization or about 15 months at 25% utilization, before staffing, cooling and networking. Actual costs depend on use and operating expenses.
Source: ThorstenMeyerAI.com
