🔍 Read the full analysis: How Close Are Open-Weight Models To On-Prem AI? on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
IBM CEO Arvind Krishna said open-weight models can cost 10% to 25% as much as frontier models while delivering about 90% of their effectiveness. Artificial Analysis data read October 26 puts the best downloadable model at 80% of the top proprietary model, with a higher-scoring model’s weights not yet released. On-prem economics also depend on utilization, licensing and the staff needed to operate the systems.
IBM CEO Arvind Krishna said on October 21 that open-weight AI models can cost 10% to 25% as much as frontier systems while being “90% as effective,” making on-premise deployment attractive for some enterprise work. An Artificial Analysis comparison read October 26 puts the leading downloadable model at 80% of the top proprietary model on its index, while a closer-scoring contender has not yet released its weights.
Krishna made the comparison during IBM’s third-quarter earnings-call Q&A. He said open models can be run on a company’s own infrastructure or private cloud, and that a large enterprise might begin with two or three racks. He pointed to routine tasks such as summarizing email, reviewing contracts and answering customer calls as work that may not need the highest-performing model. IBM has a commercial interest in this shift: it sells on-premise infrastructure, including its Spyre accelerator cards.
The Artificial Analysis Intelligence Index lists Xiaomi’s MiMo-V2.6-Pro at 46.3 and proprietary Claude Opus 5.5 at 57.6. That makes the downloadable model about 80% of the leader’s score, up from 51% at the end of 2024, according to the source’s comparison using the current index version. The gap is not erased: the score difference is 11.3 points, compared with 7.5 points at the end of 2024.
Kimi K3.1 Pro scores 55.1, or about 96% of Claude Opus 5.5 on that index. But Moonshot had only promised to release its weights by November 4. Until those weights are available, Kimi does not qualify as an open-weight model in the comparison. Scores also do not by themselves show whether a model will meet a specific company’s accuracy, privacy, latency or reliability requirements.
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 comparison matters because model capability is only one part of the decision. For businesses handling sensitive records, on-premise or private-cloud deployment can offer more control over where data and logs are processed. Krishna said banks and traders want sovereign AI, and argued that open-weight licences can permit deployment on company infrastructure. Those benefits depend on the specific model’s licence and the organization’s security setup; downloadable weights do not automatically remove every data or compliance risk.
The hardware calculation can also change with usage. The source puts an eight-GPU H200 server at a list price of $172,643 and AWS rental of the same GPU class at $63.30 an hour. On those figures, purchase cost equals about 114 days of rental at continuous full use, or about 15 months at 25% use, before staffing, cooling and networking. These are simplified comparisons, not full operating-cost estimates. A lightly used server may not deliver the savings implied by a per-token price comparison.
Organizations also need people to run the systems. Krishna acknowledged shortages in kernel and networking expertise, saying the skills gap remains outside the most sophisticated enterprises. That makes the decision a trade-off between rented, managed access and the cost and control of operating infrastructure—not simply a choice between expensive and cheap tokens.
Popular right now
TCL 85 Designer Series QD-Mini LED TV (Model 85A400U) with Google Smart TV
Ninja DualZone Smart Air Fryer, 10 Qt, 6-in-1, DZ950
Dyson Cyclone V10 Origin Cordless Vacuum
As an affiliate, we earn on qualifying purchases.
How the Performance Gap Changed
Krishna described open models as having moved from far behind proprietary leaders in late 2024 to much closer by late 2025. His figures were examples, rather than results from a named benchmark. Artificial Analysis provides one independent comparison in the supplied material: it applies the same ten evaluations across models and labels whether weights are downloadable. On that index, open models rose from 51% of the proprietary leader’s score at the end of 2024 to 74% at the end of 2025 and 80% in October 2026.
That relative progress does not mean proprietary scores stood still. The source reports the top proprietary score rising from 15.2 at the end of 2024 to 57.6 in October 2026, while the best available open-weight score rose from 7.7 to 46.3. Model rankings and scores depend on the benchmark and its version, so these figures describe this particular index, not every task or deployment.
Price comparisons need similar care. The source says hosted DeepSeek V4.1 Flash 1017, whose weights carry an MIT licence, is listed at 3% to 6% of Claude Opus 5.5’s price. It also cites closed GPT-6 Luna at $0.10 per million input tokens and $0.50 per million output tokens. Those examples show that a low token price is not exclusive to open models, and hosted pricing is not the same as the cost of buying and operating on-prem hardware. IBM also said its llm-d project can run large models at five to ten times lower cost; the supplied material does not provide an independent measurement of that claim.
Costs, Licences and Weights
Several factors remain unsettled. Kimi K3.1 Pro’s weights were promised for November 4, but the supplied material does not confirm their release or identify a published licence. Its index score therefore cannot yet establish that enterprises can download and use it on their own systems. Licences vary across models: some are MIT-licensed, while others use custom or commercial terms that may restrict use.
The hardware payback figures rely on list prices and assumed utilization. They do not include the full costs of personnel, cooling, networking, maintenance or financing, and actual cloud rates may differ. Nor does the index establish that a model scoring close to a proprietary leader will perform equally well on a company’s particular workload. Nvidia had not commented on the reported GeForce firmware-limit workaround described in the source; its terms of use prohibit circumventing technical limitations, according to that material.
Watch Model Releases and Deployment Costs
The next near-term test is whether Moonshot releases Kimi K3.1 Pro’s weights by its stated November 4 date and publishes terms that permit the intended uses. Updated independent scores and licence details will help clarify whether its reported 55.1 score translates into a practical on-prem option.
For enterprises, the next step is workload-specific evaluation: compare model performance and hosted prices against the expected utilization and full operating costs of owned hardware. IBM’s sales and deployment figures may provide evidence of demand, but the supplied material does not give a verified customer-cost study showing that the cited savings apply broadly.
Key Questions
Are open-weight models already 90% as capable as proprietary models?
Not across every model or task. On the Artificial Analysis comparison cited here, the best downloadable model scored 80% of the top proprietary model on October 26, 2026. IBM CEO Arvind Krishna’s 90% figure is his broader assessment, not a score from that index.
Can companies run open-weight models entirely on their own servers?
Some licences permit on-premise use, but permission depends on the individual model’s terms. Companies also need suitable hardware, security controls and staff to operate the system. Downloadable weights alone do not settle those requirements.
Does buying an H200 server beat renting cloud GPUs?
It depends on utilization and total operating costs. The cited estimate says an $172,643 eight-GPU server equals about 114 days of AWS rental at full use or 15 months at 25% use, before staffing, cooling and networking. These are list-price calculations, not a complete cost comparison.
Is Kimi K3.1 Pro an open-weight model now?
The supplied source says Moonshot had promised to release its weights by November 4. Until the weights are released and their licence is known, it should not be counted as an available open-weight model.
Source: ThorstenMeyerAI.com
