🔍 Read the full analysis: IBM’s CEO Says Open-Weight AI Is Nearing An On-Prem Moment 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 AI models can approach frontier systems at a fraction of their token cost, making on-premises deployment attractive to some enterprises. A benchmark cited in the source puts the best downloadable model at 80% of the leading proprietary model’s score; actual savings depend on licensing, hardware use and staffing.
IBM CEO Arvind Krishna said on the company’s October 21 earnings call that open-weight AI models can deliver about 90% of frontier-model performance at 10% to 25% of the token cost, bolstering the case for companies to run models on their own infrastructure. A benchmark cited in the source material puts the top downloadable model at 80% of the leading proprietary model’s score as of October 26, while the cost and staffing case for on-premises deployment depends on how heavily systems are used.
Krishna made the comparison during the Q&A, describing improvements in open models over the preceding two years. He said that in late 2024 a leading model might score 70 on a broad benchmark while an open model scored 20 or 30; in late 2025, he said, the comparison had moved to 70 against 50. His conclusion was that some businesses may accept lower performance for a lower price. The figures were illustrative, not tied to a named benchmark in his remarks.
The source’s comparison with the Artificial Analysis Intelligence Index shows the best downloadable model, Xiaomi’s MiMo-V2.6-Pro, scoring 46.3, against 57.6 for the leading proprietary model, Claude Opus 5.5. That is about 80% of the leader’s score. The source reports the ratio at 51% at the end of 2024, indicating that open models have narrowed the relative gap, though the raw score difference in the cited comparison is now 11.3 points, versus 7.5 at the end of 2024.
Krishna also said banks and traders may value running AI “in a way that is sovereign,” with sensitive data and logs kept under their control. Open-weight models, he said, provide permission and intellectual-property rights to run them on premises or in private clouds. He acknowledged that many organizations lack the kernel and networking engineers needed to operate these systems, describing that capability gap as a continuing barrier outside the most sophisticated enterprises.
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
The proposal matters because it could change how companies balance AI performance, operating costs and data control. An on-premises system may appeal to organizations handling sensitive information or seeking more direct control over where data and logs are processed. But downloadable weights do not automatically make a deployment cheaper: companies must account for the hardware, its utilization, power, cooling, networking and staff.
The source gives an illustrative hardware comparison: an eight-GPU H200 server is listed from $172,643, while renting the same GPU class on AWS costs $63.30 an hour. Its calculation estimates that purchase costs could be recovered in about 114 days at full utilization or about 15 months at 25% utilization, before staffing and other operating expenses. Those estimates depend on the listed prices and assumptions; they are not a universal break-even point.
Model prices also vary. 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. It also cites OpenAI’s closed GPT-6 Luna at $0.10 per million input tokens and $0.50 per million output tokens. These examples show why the token-cost comparison alone cannot settle whether buying and operating hardware is worthwhile.
Popular right now
TCL 85 Designer Series QD-Mini LED TV (Model 85A400U) with Google Smart TV
Shark AI Ultra Robot Vacuum with Self-Empty Base
Bose QuietComfort Headphones (2nd Gen), Wireless Noise-Cancelling Over-Ear Headphones, Eucalyptus Green
As an affiliate, we earn on qualifying purchases.
From Model Scores to Enterprise Racks
Krishna’s argument combines two developments: the reported improvement in open-weight model scores and the prospect of deploying those models on infrastructure controlled by the customer. Open-weight systems make model parameters available for use, subject to their licences. That is different from a hosted service, where a provider operates the model, and it does not mean every model can be used without conditions.
The source’s Artificial Analysis figures show the strongest downloadable model at 7.7 compared with a proprietary leader at 15.2 at the end of 2024, or 51% of the leader. It reports 22.4 versus 30.4 at the end of 2025, and 33.7 versus 49.6 at the end of June 2026. For October 26, 2026, the figures are 46.3 versus 57.6. They are results from the index version and models identified in the source; Krishna’s illustrative 70, 50 and 90 figures should not be treated as those index scores.
A possible exception in the latest comparison is Moonshot’s Kimi K3.1 Pro, which the source says scored 55.1, or 96% of Claude Opus 5.5. But Moonshot had only promised to release its weights by November 4. Until they are available, the model is not open-weight under the source’s comparison. Licence terms also differ across models, so a high score alone does not establish that a company can use a model for its intended purpose.
Benchmark Gaps and Deployment Costs
The benchmark comparison does not establish that open models are 90% as effective across all tasks. The source reports a score ratio of about 80% for the top downloadable model and says the figures Krishna gave were illustrative, not results from a named benchmark. Benchmark scores also do not predict performance on every company’s workloads.
Actual total costs remain uncertain. The hardware comparison excludes staffing, cooling and networking, and assumes particular levels of use. The source does not provide a full enterprise cost analysis or independently verify IBM’s separate claim that its llm-d project can run large models at five to 10 times lower cost. It also notes that IBM sells on-premises hardware, making it an interested party in the shift Krishna described.
Kimi K3.1 Pro’s weights and licence had not been released in the source’s account, despite a promised date of November 4. Until those details are available, its score cannot be treated as evidence that customers can already download and deploy the model. The source also does not say how many enterprises are preparing deployments or how soon broad adoption might occur.
Model Releases and Enterprise Pilots
The next near-term check is whether Moonshot releases Kimi K3.1 Pro’s weights by its stated November 4 target and publishes licence terms. That would determine whether the model belongs in a comparison of downloadable models and clarify what users may do with it.
For enterprise buyers, the next evidence will come from deployments that disclose workload performance, hardware utilization and full operating costs. Companies weighing the approach will also need to assess whether they have the engineering staff and network capacity Krishna identified as requirements. IBM has positioned its hardware and software around this opportunity, but the earnings-call remarks and benchmark data do not establish how quickly customers will adopt it.
Key Questions
What did IBM’s CEO say about open-weight AI?
Arvind Krishna said on IBM’s October 21, 2026, earnings call that open-weight models can cost 10% to 25% as much per token as frontier models while being about 90% as effective. He presented that as an illustrative comparison, not a result from a named benchmark.
How close are open models to proprietary models in the cited benchmark?
The source reports that MiMo-V2.6-Pro scored 46.3 on the Artificial Analysis Intelligence Index, compared with 57.6 for Claude Opus 5.5, as of October 26, 2026. That puts the downloadable model at about 80% of the proprietary model’s score in this comparison.
Does running AI on-premises automatically cost less?
No. Savings depend on purchase or rental prices, system utilization and expenses such as staffing, cooling and networking. A lightly used server can take much longer to offset its purchase price than one operating near full capacity.
Is Kimi K3.1 Pro available as an open-weight model?
Not according to the source material as of October 26, 2026. It says Moonshot had promised to release the weights by November 4, but they had not yet shipped and no licence had been published.
Source: ThorstenMeyerAI.com
