🔍 Read the full analysis: IBM’s AI Pitch: Run Models Yourself As Open Weights Advance on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
IBM CEO Arvind Krishna argued on the company’s October 21 earnings call that open-weight models can deliver about 90% of frontier-model performance at 10% to 25% of the token cost, and can run on enterprise infrastructure. An October 26 Artificial Analysis comparison puts the best available open model at about 80% of the top proprietary model; the economics of running models in-house also depend on utilization, staffing and hardware costs.
IBM CEO Arvind Krishna said on the company’s October 21 earnings call that open-weight AI models can cost 10% to 25% as much per token as leading proprietary models while being “90% as effective,” making them a potential fit for companies that want to run AI on their own infrastructure. An independent model comparison available October 26 put the strongest downloadable model at 80% of the top proprietary model’s score, highlighting both the progress and the remaining gap behind IBM’s pitch.
Krishna made the remarks during the earnings-call Q&A, describing a narrowing performance gap. He said that in late 2024 a leading model might score 70 on a broad benchmark while an open model scored 20 or 30; by late 2025, he said, the comparison was about 70 to 50. He then posed the cost trade-off: “Would you like something that’s 10% of the cost if it’s 90% as good as the best, or do you want to pay that 10 times higher cost?” Those figures were his illustrations, not results from a named benchmark.
Krishna said open-weight models may be sufficient for routine work such as summarizing emails, reviewing contracts and answering customer calls. He also pointed to data control: banks and trading firms may want to keep underlying data and logs within systems they control. Open weights can give customers permission to run a model on their own hardware or private cloud, but rights and restrictions depend on each model’s licence.
In the Artificial Analysis Intelligence Index, read October 26, MiMo-V2.6-Pro scored 46.3, compared with 57.6 for Claude Opus 5.5. The open model’s score was about 80% of the proprietary leader’s. Kimi K3.1 Pro scored 55.1, about 96% of the leader, but its weights had not yet been released; its developer had promised them by November 4. Until release, it does not count as a downloadable open-weight model in this 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 of Running Models In-House
For companies choosing AI systems, the decision is not simply whether a model is open or closed. It is a calculation involving model quality, token prices, data governance and infrastructure use. A less capable model may be adequate for common tasks, while a more demanding task could still require a leading proprietary system. The benchmark comparison gives buyers one reference point, but it does not establish that every model performs equally well across their own workloads.
Owning hardware can lower costs for firms that keep it busy, but the source’s price examples show how sensitive the calculation is to utilization. An eight-GPU H200 server was listed at $172,643, while renting the same GPU class on AWS was listed at $63.30 an hour. At full use, the purchase cost would equal about 114 days of rental; at 25% use, the comparison stretches to roughly 15 months. Those estimates exclude staff, cooling and networking, so they are not a complete cost-of-ownership analysis.
Krishna acknowledged that operating systems is a constraint, citing shortages of kernel and networking engineers. That matters because the hardware purchase alone does not deliver a working AI service. Companies must also plan for deployment, maintenance and the expertise needed to keep infrastructure productive. The pitch is most relevant to organizations with enough recurring demand and technical capacity to support those costs.
Products Worth Considering
DJI Osmo Mobile 8 Standard Combo Phone Gimbal Stabilizer
LG UltraGear 32GX850A-B 32-Inch 4K UHD Glossy OLED Gaming Monitor
Samsung 27 Odyssey OLED G5 (G50SF) QHD QD-OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
From Benchmark Gains to Enterprise Deployment
The source’s benchmark history shows the open-model share of the leading proprietary score rising from 51% at the end of 2024 to 74% at the end of 2025, then 68% at the end of June 2026 and 80% on October 26. The comparison uses Artificial Analysis’s model index and its open-versus-proprietary chart. It supports the broad direction of Krishna’s argument, but not a universal claim that open models are 90% as capable: the exact result varies by model and benchmark, and the best available open model remained below the leader.
Price comparisons also depend on how a company accesses a model. The source says hosted DeepSeek V4.1 Flash 1017, whose weights are MIT-licensed, costs 3% to 6% of Claude Opus 5.5’s listed price. But it also cites OpenAI’s closed GPT-6 Luna at $0.10 per million input tokens and $0.50 per million output tokens. Those examples show that lower prices do not follow automatically from open weights, and that token pricing for hosted services is different from the cost of buying and operating hardware.
IBM has a commercial interest in greater on-premises AI deployment. Krishna said the company expects demand to support sales of its Spyre cards; the source also reports that IBM’s mainframe revenue fell 20% in the quarter. IBM said its llm-d project can run large models at five to 10 times lower cost, but that is a company claim, not an independently verified measurement in the material provided. The article’s separate report on a Linux module that raises performance limits on some GeForce cards is not evidence about enterprise model economics; it concerns a distinct, enthusiast-focused hardware issue.
Performance, Licensing and Cost Caveats
The benchmark figures do not settle how an open model will perform on a particular company’s tasks. Krishna’s 90% figure was not linked to a named evaluation, while the Artificial Analysis scores summarize results across ten evaluations and may not predict performance for a specific deployment. It is also unclear whether Kimi K3.1 Pro’s promised weights will arrive by November 4, or what licence will govern them; the source says no licence had been published as of its reporting date.
Licence terms vary, and a high leaderboard score does not by itself establish that a model can be used commercially without conditions. Hardware comparisons also leave out operational expenses and depend heavily on actual utilization. The material does not provide a full, independently audited comparison of total costs for IBM’s proposed on-prem approach, competing cloud services and model providers.
The Next Test Is Deployment
The near-term test is whether promised model weights, including Kimi K3.1 Pro’s, are released on schedule and under terms that allow the uses buyers expect. Enterprises considering deployment will need to check model licences, evaluate performance on their own tasks and estimate demand before comparing cloud bills with hardware purchases.
IBM’s claims about on-premise savings and its llm-d project will need independent cost measurements to show how they hold up across real workloads. The source provides no confirmed timetable for such verification. For now, the evidence points to a narrowing capability gap and potentially lower prices in some cases, while the cost advantage of running models yourself remains dependent on workload, utilization and in-house expertise.
Key Questions
What did IBM’s CEO say about open-weight models?
Arvind Krishna said on IBM’s October 21 earnings call that open-weight models could cost 10% to 25% as much per token as leading proprietary models while being about 90% as effective. He presented the figures as an argument for their use, not as a result from a named benchmark.
How close are the best available open models to proprietary leaders?
In the Artificial Analysis comparison read October 26, MiMo-V2.6-Pro scored 46.3, about 80% of Claude Opus 5.5’s 57.6. Kimi K3.1 Pro scored 55.1, but its weights had not been released at the time.
Does running an open model on your own servers always cost less?
No. The economics depend on hardware utilization, staffing and operating costs, as well as the model and workload. The cited server-versus-rental comparison excludes expenses such as cooling, networking and staff.
Do open weights mean a model has no usage restrictions?
No. Model licences differ, including commercial and community licences with varying conditions. Buyers need to review the specific licence; downloadable weights do not automatically mean unrestricted use.
Source: ThorstenMeyerAI.com
