🔍 Read the full analysis: Are Open-Weight Models Close Enough For On-Prem AI? 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 deliver about 90% of frontier-model effectiveness at 10% to 25% of the token cost, pitching on-premises deployment as an option for large enterprises. An Artificial Analysis index read Oct. 26 puts the best currently downloadable model at 80% of the proprietary leader; hardware economics and the engineering needed to run models locally remain key uncertainties.
IBM CEO Arvind Krishna said on the company’s Oct. 21 earnings call that open-weight AI models can be about 90% as effective as leading proprietary models at 10% to 25% of their token cost, making on-premises deployment worth considering for some enterprises. An Artificial Analysis comparison read Oct. 26 shows the strongest model with downloadable weights at 80% of the proprietary leader’s score, while the cost case for buying servers depends heavily on how intensively they are used.
Krishna said large enterprises could begin deploying the systems with two or three racks. He cited routine tasks such as summarizing email, reviewing contracts, answering customer calls and drafting messages as examples that may not require the most capable available model. He also argued that running models on a company’s own infrastructure or private cloud can help organizations retain control over data and logs. These are IBM’s arguments for adoption, not guarantees that every workload will be cheaper or more secure on-premises.
The Artificial Analysis Intelligence Index compares models across ten evaluations. Its Oct. 26 listing gives MiMo-V2.6-Pro, the highest-scoring model identified in the source as currently downloadable, a score of 46.3, versus 57.6 for Claude Opus 5.5. That places the open model at about 80% of the proprietary model’s score. The source’s historical comparison puts the 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, listed at 55.1, or 96% of Claude Opus 5.5’s score. But the source says Moonshot had promised to release its weights by Nov. 4; until those weights are available, it is not a downloadable open-weight option. The score comparison also does not by itself establish performance on a particular company’s tasks.
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 Depends on Local Use
The choice is not simply between a low-cost model and an expensive one. Companies considering on-prem AI must weigh model quality against hardware purchase, operating costs and utilization. The source gives a list price of $172,643 for an eight-GPU H200 server and an AWS rental rate of $63.30 an hour for the same GPU class. On those figures, buying could match rental costs in about 114 days at full utilization, but about 15 months at 25% utilization. Those estimates exclude staffing, cooling and networking.
That utilization gap can change the economics for organizations with uneven or modest workloads. Renting can avoid a large upfront purchase and leave infrastructure management to a provider; owned hardware may make more sense when it is kept busy. IBM’s own remarks acknowledged a staffing challenge: Krishna asked where companies would find kernel and networking engineers, and said that capability gap remains for all but the most sophisticated enterprises.
There is also a meaningful trade-off between data control and operational responsibility. Running weights locally can give an organization more direct control over where workloads and logs are processed, but it also places more of the deployment burden on the organization. The article’s comparison supports treating open models as a growing option—not as proof that every enterprise can replace frontier services at a tenth of the total cost.
Products Worth Considering
Jackery Explorer 300 Portable Power Station, 292Wh
Amazon Kindle Paperwhite Signature Edition (12th Generation, 2024), 32 GB, Metallic Black
Apple iPad Air 11-inch (M2), 512GB, Wi-Fi, Space Gray
As an affiliate, we earn on qualifying purchases.
How the Model Gap Has Shifted
Krishna’s “90%” comparison was a broad characterization on an earnings-call Q&A, not a result tied to a named benchmark. He described open models as having moved from scores of roughly 20 or 30 against a leading model at 70 in late 2024 to about 50 against 70 in late 2025. The Artificial Analysis figures offer a separate way to track the trend, but they use that index’s evaluations and model entries rather than validating Krishna’s specific example.
On the index figures cited in the source, the share rose from 51% at the end of 2024 to 80% on Oct. 26, 2026. The absolute score gap, however, is reported as 7.5 points then and 11.3 points now, because proprietary models also improved. Rankings and comparisons depend on which models are included and the index version; a general score does not establish results for a company’s own applications.
“Open-weight” also does not mean every model has the same terms of use. The source identifies MiMo-V2.6-Pro and DeepSeek V4.1 Flash 1017 as MIT-licensed, while other models use their own commercial, community or conditional licences. Organizations need to check the relevant licence and intended use. The article also reports that DeepSeek’s hosted V4.1 Flash 1017 costs 3% to 6% of Claude Opus 5.5’s listed price, but that hosted-service comparison is not the same as the full cost of operating a local deployment.
Benchmark Scores Are Not Total Cost
The benchmark establishes neither that open models are 90% as effective across enterprise work nor that on-prem deployment costs 10% of a proprietary service in total. The figures compare scores and listed token or rental prices; they do not include every organization’s staffing, energy, cooling, network, security and maintenance expenses. The source does not provide a workload-specific trial or a complete cost study.
The status of Kimi K3.1 Pro’s weights was also pending in the source material: Moonshot had promised them by Nov. 4, but the release and licence were not confirmed there. The source says no licence had been published. Without both the weights and applicable terms, the model cannot be counted as an available open-weight deployment on the basis of its score alone.
IBM has a commercial interest in on-prem infrastructure and sells hardware, including Spyre cards. The source also reports that IBM’s mainframe revenue fell 20% in the quarter, citing The Register. That context does not disprove Krishna’s claims, but it is relevant when weighing the company’s projections. IBM’s claim that its llm-d project can run large models at five to ten times lower cost is likewise not independently measured in the supplied material.
Watch for Weights and Workload Tests
The next concrete check identified in the source is whether Moonshot releases Kimi K3.1 Pro’s weights by its promised Nov. 4 date, and what licence accompanies them. A release would make its benchmark score more relevant to organizations evaluating downloadable models, though it would not answer questions about hardware requirements or performance on specific tasks.
For enterprises weighing a deployment, the practical next step is to compare representative workloads on candidate models and include the costs of hardware, cloud alternatives and specialist staff. The source provides no confirmed industry-wide adoption figure or date by which on-prem AI will become the default. The economic case will remain dependent on each organization’s workload, utilization and licence requirements.
Key Questions
What did IBM’s CEO say about open-weight models?
Arvind Krishna said open-weight models can be about 90% as effective as frontier models at 10% to 25% of their token cost. He presented this as a general comparison on IBM’s earnings call, not as a result from a named benchmark.
How close are the best downloadable models to proprietary leaders?
In the Artificial Analysis data cited as read on Oct. 26, 2026, MiMo-V2.6-Pro scored 46.3, compared with 57.6 for Claude Opus 5.5—about 80% of the proprietary model’s score on that index.
Is Kimi K3.1 Pro already available as an open-weight model?
Not according to the supplied source. Moonshot had promised to release the weights by Nov. 4, but the source says they had not yet shipped and no licence had been published.
Does on-prem AI automatically cost less than cloud AI?
No. The cited server and rental figures suggest the purchase can pay off faster with high utilization, but the estimates exclude staff, cooling and networking. Workload volume and operating costs affect the comparison.
What should a company check before using an open-weight model?
It should test the model on its own tasks, estimate full deployment and operating costs, and review the model’s licence and usage conditions. A benchmark score or low hosted token price alone does not establish the total cost or suitability of a local deployment.
Source: ThorstenMeyerAI.com
