🔍 Read the full analysis: IBM’s AI Outlook: Open-Weight Models Near, On-Prem Depends On Volume on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
IBM CEO Arvind Krishna said on the company’s October 21 earnings call that open-weight models can deliver about 90% of frontier-model quality for 10% to 25% of the token cost. An independent benchmark puts the strongest currently downloadable model at 80% of the leading proprietary score; whether running models on-premises saves money depends on utilization, staffing and infrastructure.
IBM CEO Arvind Krishna said on the company’s October 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 attractive for some enterprises. An Artificial Analysis index read October 26 puts its best currently downloadable model at 80% of the top proprietary score, while cost comparisons show that buying hardware pays off only with sustained use and does not account for operating expenses.
Krishna made the comparison during the call’s question-and-answer session. He described open models as improving against proprietary systems and argued that many routine business tasks—such as summarizing email, reviewing contracts and answering customer calls—do not require the best available model. The figures were his characterization, not results from a named benchmark. IBM also has a commercial interest in on-premises AI, including sales of its hardware and infrastructure products.
On Artificial Analysis’s Intelligence Index, the highest-scoring downloadable model listed on October 26 was Xiaomi’s MiMo-V2.6-Pro, at 46.3. The top proprietary model, Claude Opus 5.5, scored 57.6, putting the open model at about 80% of the leader’s score. The source’s comparison shows the share was 51% at the end of 2024 and 74% at the end of 2025. These results support a narrowing relative gap, but do not verify Krishna’s 90% figure across tasks.
A possible exception is Kimi K3.1 Pro, which scored 55.1 on the index, or about 96% of the proprietary leader. But its weights had not been released as of October 26: Moonshot had promised them by November 4. Until downloadable weights are available, it is not an open-weight model in the practical sense relevant to an organization seeking to run it on its own infrastructure.
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
When On-Premises AI Can Pay
The appeal is not only a lower inference bill. Krishna said banks and trading firms may want to run AI in a sovereign environment, limiting where underlying data and logs are held. Open weights can give organizations more control over deployment, subject to the model’s licence, security practices and the expertise required to operate it. That may matter to firms with strict data-handling rules or large, predictable workloads.
But a lower price per token does not automatically translate into lower total cost. The source’s illustrative comparison lists an eight-GPU H200 server at $172,643 and rental of the same GPU class on AWS at $63.30 an hour. On those listed prices, the purchase cost equals roughly 114 days of continuous rental. At 25% utilization, the payback period stretches to about 15 months, before adding staff, cooling and networking. These are simple comparisons, not a full ownership-cost analysis, and real prices and workloads vary.
Krishna acknowledged an implementation constraint: enterprises need engineers who can handle kernels and networking. He said that talent gap remains outside the most sophisticated organizations. That makes the decision less about model scores alone and more about workload volume, available staff, infrastructure costs and how much control the organization needs.
Products Worth Considering
Bose QuietComfort Ultra Earbuds (2nd Gen), Deep Plum
Amazon Kindle Scribe Colorsoft 64GB (1st Generation)
Logitech MX Keys S Wireless Keyboard, Graphite
As an affiliate, we earn on qualifying purchases.
The Benchmarks Behind the Claim
Krishna’s “90% as effective” phrase was an illustrative comparison on an earnings call, rather than a claim tied to a particular evaluation. Artificial Analysis provides one independent comparison by running models through a consistent set of ten evaluations and distinguishing models with downloadable weights. Its figures show the relative share of the leading proprietary score rising from 51% at the end of 2024 to 80% in October 2026. The absolute score gap, however, was 7.5 points in the end-2024 comparison and 11.3 points in October 2026, because proprietary scores also rose.
Price comparisons depend on how a model is accessed. The source says DeepSeek V4.1 Flash 1017, whose weights carry an MIT licence, is priced by its host at 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. Those examples show why “open” and “cheap” are not interchangeable categories: hosted prices, model capability, licence terms and the cost of running hardware all matter.
Licences differ even among models with available weights. The source identifies MiMo-V2.6-Pro and DeepSeek V4.1 Flash 1017 as MIT-licensed, while some other models use custom commercial or community terms. Organizations need to check the applicable licence rather than infer commercial rights from a model’s presence on a public leaderboard.
Weights, Licences and Operating Costs
The comparison does not establish that open-weight models match proprietary systems across every task. Artificial Analysis’s index is one standardized measure, and the available source does not provide enough detail to determine how its scores translate to each organization’s specific workloads. Krishna’s estimate remains an IBM executive’s claim, not a universal performance or savings guarantee.
Kimi K3.1 Pro’s weights were not available by the October 26 benchmark date, and its licence had not been published in the source material. Its score cannot establish that enterprises can already download and operate it. The on-premises cost figures also omit factors such as power, cooling, networking, staffing, maintenance and utilization patterns beyond the simplified examples. IBM’s separate statement that its llm-d project can run large models at five to ten times lower cost is likewise an IBM claim; the source does not provide an independent measurement to verify it.
Model Releases and Enterprise Pilots
The next near-term test is whether Moonshot releases Kimi K3.1 Pro’s weights by its promised November 4 date, and what licence terms accompany them. A release would change whether the model can be counted among downloadable options, but organizations would still need to evaluate hardware requirements and deployment rights.
For buyers, the practical next step is to test candidate models against their own tasks and compare the full cost of hosted inference with the cost of running and maintaining infrastructure. Further benchmark updates may shift the performance comparison, while IBM’s claims about on-premises demand and its software’s cost savings will require evidence from deployments to establish how broadly they hold.
Key Questions
What did IBM’s CEO say about open-weight models?
Arvind Krishna said on October 21 that open-weight models could be about 90% as effective as frontier models at 10% to 25% of their token cost. He did not tie the 90% figure to a named benchmark in the remarks described by the source.
How close is the best downloadable model to the top proprietary model?
On the Artificial Analysis index read October 26, MiMo-V2.6-Pro scored 46.3, compared with 57.6 for Claude Opus 5.5. That is about 80% of the proprietary model’s score on that index.
Is Kimi K3.1 Pro already an open-weight model?
Not according to the source’s October 26 status. It says Moonshot had promised to release the weights by November 4, but they had not yet shipped and no licence had been published.
Does buying AI hardware always cost less than renting it?
No. The source’s comparison estimates that an eight-GPU H200 server’s listed purchase price equals about 114 days of continuous rental at the cited AWS rate. At 25% utilization, the estimated payback becomes about 15 months, before staffing and other operating costs.
Why might an enterprise choose on-premises AI?
Some organizations may want greater control over where data and logs are handled, as well as predictable costs for heavily used workloads. The trade-off is the need to fund hardware and employ staff able to operate it.
Source: ThorstenMeyerAI.com
