🔍 Read the full analysis: Open-Weight Models Are Getting Close, IBM’s CEO Says on ThorstenMeyerAI.com
Get hardware and tech essentials 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 AI models can deliver about 90% of frontier-model performance at 10% to 25% of the token cost, and can be run on enterprise infrastructure. An Artificial Analysis comparison puts the leading downloadable model at 80% of the top proprietary model as of October 26, while a higher-scoring model’s weights have not yet been released.
IBM CEO Arvind Krishna said open-weight AI models are becoming capable enough for many enterprise tasks while costing as little as 10% to 25% of frontier models’ token fees, making it worth considering whether to run them on company infrastructure. A benchmark comparison available October 26 put the leading downloadable model at 80% of the top proprietary model’s score, short of Krishna’s 90% estimate; the comparison also involves a potential contender whose weights have not yet been released.
Krishna made the comparison during the question-and-answer session of IBM’s October 21 earnings call. He described open models as “90% as effective” while costing “25% to 10% of the token cost of the frontier models.” He said larger enterprises could begin on-premises deployments with two or three racks. His examples of tasks that may not require the most capable model included summarizing email, examining contracts, answering customer calls and composing messages.
The independent comparison cited in the source material is Artificial Analysis’s Intelligence Index, which applies the same set of ten evaluations across models and identifies whether weights can be downloaded. In the index as read on October 26, Xiaomi’s MiMo-V2.6-Pro scored 46.3, compared with 57.6 for Anthropic’s Claude Opus 5.5. That puts the open model at about 80% of the proprietary leader’s score. The measure is a benchmark result, not a universal measure of performance across every task or deployment.
Krishna’s figure is also a broad characterization rather than a result tied to a named benchmark. He illustrated the change over time by saying open models had moved from scores of 20 or 30 against a leading model scoring 70 in late 2024, to 50 against 70 in late 2025. The Artificial Analysis figures show a different but similar direction: its comparison puts the open-to-proprietary score ratio at 51% at the end of 2024 and 80% in October 2026. The comparison does not show that open models now match the frontier leader.
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 Enterprise AI
The choice is not simply between a powerful model and a cheaper one. Companies must compare model quality, token fees and deployment costs, including hardware, staffing, power and network infrastructure. Krishna’s argument is that some routine workloads may be good enough on a less costly model, while sensitive data can remain within a company’s own systems or private cloud.
That case has particular relevance for organizations that handle confidential financial or customer information. Krishna said banks and traders want to run AI “in a way that is sovereign,” so that outside parties cannot access underlying data or logs. Open weights can give organizations permission to run models on their own infrastructure, depending on the model’s licence. They do not, by themselves, guarantee privacy, security or regulatory compliance.
The financial comparison depends heavily on utilization. The source material lists an eight-GPU H200 server at about $172,643 and AWS rental of the same GPU class at $63.30 an hour. Its calculation estimates a purchase could match rental costs in about 114 days at full utilization, but about 15 months at 25% utilization, before staffing, cooling and networking. Those estimates are not a complete cost-of-ownership analysis; they illustrate why an on-premises investment can look different for a company with steady demand than for one with intermittent workloads.
There is also a commercial interest behind IBM’s argument. IBM sells hardware and services that could benefit from greater on-premises AI use. The source material says Krishna expects demand to support sales of IBM’s Spyre cards, and reports that IBM mainframe revenue fell 20% in the quarter. That does not invalidate his comments, but it is relevant context for readers weighing the claims.
Products Worth Considering
Apple MacBook Pro 14.2-inch with M5 Max, 18-Core CPU, 20-Core GPU, 48GB Memory, 2TB SSD, Space Black
Apple MacBook Pro 16-inch Laptop with M5 Pro, 18-Core CPU, 20-Core GPU, 48GB Memory, 2TB SSD, Space Black
Apple MacBook Pro 16.2-inch with M5 Pro, 18-Core CPU, 20-Core GPU, 48GB Memory, 1TB SSD, Nano-Texture Display, Space Black
As an affiliate, we earn on qualifying purchases.
From Open-Model Gap to Deployment Costs
The benchmark trend cited by the source shows the leading downloadable model’s score rising relative to the leading proprietary model over the period examined. At the end of 2024, Artificial Analysis listed o1 at 15.2 and Qwen2.5 72B at 7.7, or 51% of the proprietary score. By the end of 2025, the comparison was GPT-5.2 (xhigh) at 30.4 and MiMo-V2-Flash at 22.4, or 74%. The October 2026 comparison puts the ratio at 80%.
One possible exception is Kimi K3.1 Pro, with a listed score of 55.1, roughly 96% of Claude Opus 5.5’s 57.6. But the source says Moonshot had only promised to release the weights “by Nov 4.” Until the weights are available, it is not an open-weight model that enterprises can download and run. The source also says no licence had been published, leaving the terms of any release unresolved.
Licensing and hosting arrangements differ across models, so a high score alone does not establish whether a model can be used commercially or deployed as a company needs. The source lists MiMo-V2.6-Pro and DeepSeek V4.1 Flash 1017 as MIT-licensed, while other models carry custom commercial or community licences. It also cites IBM’s claim that its llm-d project can run large models at “5 to 10 times lower cost”; that figure is an IBM claim, and the source does not provide an independent measurement.
Benchmark Scores and Costs Have Limits
The benchmark comparison does not settle how models perform on every task an enterprise might run. A score on ten evaluations cannot establish whether a particular model is suitable for a company’s contracts, customer calls or internal records. The comparison also covers models and index results available on a specific date, October 26, 2026; rankings may change as models and evaluations are updated.
The projected near-match depends on Kimi K3.1 Pro’s weights being released as promised. The source says the announced date was November 4, but reports that the weights had not shipped as of October 26 and that no licence had been published. Its score therefore cannot yet be treated as a result from a downloadable model available to enterprises.
The cost figures also leave important assumptions unresolved. List prices and utilization-based payback estimates do not account for every organization’s staffing, electricity, cooling, networking, maintenance or security requirements. Krishna himself identified a shortage of kernel and network engineering expertise outside the most sophisticated enterprises. IBM’s commercial stake in on-premises deployments, and its unverified llm-d cost claim, are further reasons to distinguish company statements from independently checked measurements.
Watch for Weights, Licences and Real Costs
The next clear milestone in the source material is whether Moonshot releases Kimi K3.1 Pro’s weights by November 4, and what licence accompanies them. If the weights become available, the model could narrow the measured gap on the Artificial Analysis index, but its release terms and hardware requirements will shape whether organizations can use it in practice.
For enterprises considering on-premises AI, the practical next step is to test candidate models on their own workloads and compare the full cost of operating them with hosted alternatives. That means checking licences, hardware needs, utilization, staffing and data-handling requirements rather than relying on a single performance ratio or token price. IBM has not provided, in the material cited here, an independent accounting that verifies its broad cost and performance claims.
Benchmark standings and model prices may change as new systems appear. The confirmed picture as of October 26 is that open-weight models have gained ground on the cited index, while the leading downloadable model still trails the proprietary leader and the closest reported challenger is not yet available with released weights.
Key Questions
What did IBM CEO Arvind Krishna say about open-weight models?
Krishna said open-weight models can be about 90% as effective as frontier models while costing 10% to 25% as much per token. That is his broad claim; he did not tie the comparison to a named benchmark in the comments described here.
How close are open-weight models on the cited benchmark?
As of October 26, Artificial Analysis listed MiMo-V2.6-Pro at 46.3 and Claude Opus 5.5 at 57.6. The downloadable model’s score was about 80% of the proprietary leader’s score on that index.
Could Kimi K3.1 Pro reach the 90% mark?
The source lists Kimi K3.1 Pro at 55.1, or about 96% of Claude Opus 5.5’s score. But its weights had not been released as of October 26; Moonshot had promised them by November 4, and the source says no licence had been published.
Does running an open model on-premises always cost less?
No. Costs depend on how much the hardware is used and on staffing, power, cooling and networking. The source’s H200 server example estimates payback against rental at about 114 days at full load and 15 months at 25% utilization, before several operating costs.
What should companies check before using an open-weight model?
Companies should check the model’s licence, task performance and deployment requirements, then compare total operating costs with hosted options. A downloadable model does not by itself establish that its terms permit a particular commercial use or that it meets an organization’s security requirements.
Source: ThorstenMeyerAI.com
