🔍 Read the full analysis: IBM’s CEO Says Run AI Yourself As Open-Weight Models Draw Near 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 deliver about 90% of frontier performance at a fraction of the token cost, and that enterprises can run them on their own infrastructure. An Artificial Analysis comparison puts the leading downloadable model at 80% of the top proprietary score as of Oct. 26, 2026; a model that would score closer has not yet released its weights.
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 systems while costing a fraction as much per token, making it practical for some companies to run AI on their own infrastructure. An Artificial Analysis comparison read Oct. 26 puts the leading downloadable model at 80% of the top proprietary score, underscoring that the business case depends on the task, model access and cost of operating the hardware.
Krishna said open-weight models can cost from 10% to 25% of frontier-model token costs. He cited routine work such as summarizing email, reviewing contracts, answering customer calls and drafting messages as tasks that may not require the strongest available system. He also said a large enterprise could start with two or three racks of equipment. Those figures and judgments came from IBM’s CEO, not an independent test of enterprise workloads.
The Artificial Analysis Intelligence Index compares models across ten evaluations and identifies whether their weights can be downloaded. On the version checked Oct. 26, Xiaomi’s MiMo-V2.6-Pro scored 46.3, against 57.6 for proprietary Claude Opus 5.5. That makes the downloadable model about 80% of the leader’s score. The comparison supports a narrowing gap over time, but does not verify Krishna’s 90% figure as a general measure of model quality.
Kimi K3.1 Pro scored 55.1 on the index, which would place it near 96% of Claude Opus 5.5. However, its developer, Moonshot, had promised the weights by Nov. 4; they had not shipped at the time of the comparison. Until the weights are available, the model cannot be counted as open-weight under the index’s comparison. A strong score alone does not establish that an enterprise can download, deploy or use a model under acceptable terms.
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 Economics of Running Models
The choice between hosted AI and on-premises deployment is a cost and capacity decision, not just a comparison of model prices. A company operating its own system can gain more control over where data and logs are held, a concern Krishna raised for banks and trading firms. But it must also provide the hardware, engineering staff, networking, cooling and ongoing operations needed to keep the system useful.
The supplied price examples show why utilization matters. An eight-GPU H200 server lists from $172,643, while renting the same GPU class on AWS is quoted at $63.30 an hour. At full use, the purchase price equals roughly 114 days of rental charges; at 25% utilization, it takes about 15 months. These are simple comparisons of listed hardware and rental prices, not full cost-of-ownership estimates. They exclude staff, cooling and networking, among other operating expenses.
Token prices can also complicate the argument for buying hardware. DeepSeek’s hosted V4.1 Flash 1017 is listed at 3% to 6% of Claude Opus 5.5’s price, while OpenAI’s closed GPT-6 Luna is listed at $0.10 per million input tokens and $0.50 per million output tokens. The comparison suggests low-cost hosted options can compete with self-hosting, depending on workload and usage. It does not establish that every model, provider or deployment has comparable performance or terms.
Krishna acknowledged that building and operating an on-premises system requires skills many companies lack. He singled out kernel engineers and people who manage the underlying network as possible bottlenecks, saying the gap remains for all but the most sophisticated enterprises. The economics therefore turn on steady utilization and technical capacity, as well as the model’s price and performance.
Products Worth Considering
Samsung 32-Inch Odyssey OLED G8 (G80SH) 4K Gaming Monitor
Apple iPad Air (4th Generation, 10.9-inch, Wi-Fi, 256GB) – Space Gray
LG UltraGear 27-Inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDMI 2.1
As an affiliate, we earn on qualifying purchases.
How the Model Gap Has Shifted
Krishna described the movement in model performance as a narrowing gap: in late 2024, he said, a top model might score 70 while an open model scored 20 or 30; by late 2025, he characterized it as 70 to 50. He presented those numbers as illustrations on the earnings call, not as results from a named benchmark. Artificial Analysis offers a separate comparison: its index puts the leading open-weight model at 51% of the proprietary leader at the end of 2024, 74% at the end of 2025 and 80% on Oct. 26, 2026.
The index figures show the relative share increasing, but the absolute score gap did not shrink between the end of 2024 and Oct. 26: it was 7.5 points then and 11.3 points in the later comparison. Proprietary models also improved during that period. Scores from different dates reflect the models leading at each point and should not be read as a controlled comparison of an unchanged pair.
Access and licensing are separate from performance. MiMo-V2.6-Pro and DeepSeek V4.1 Flash 1017 are described in the source material as carrying MIT licences; other models use commercial or community terms. A model called open-weight is not automatically unrestricted for every business use. Enterprises need to check the specific licence and whether the weights are actually available. IBM also said in prepared remarks that its llm-d project can run large models at five to ten times lower cost; that is IBM’s claim, and the source material does not provide an independent measurement to verify it.
IBM has a commercial interest in the argument: Krishna said on-premises AI demand could support sales of the company’s Spyre cards. The Register, as cited in the source material, reported that IBM’s mainframe revenue fell 20% in the quarter, 18 months into the z17 cycle. That context does not disprove Krishna’s case, but it is relevant when weighing an executive’s forecast about demand for company hardware.
Performance, Costs and Access
The index measures performance across its evaluations, but it does not settle how a model will perform on a company’s own tasks, data or safety requirements. Krishna’s 90% effectiveness figure was not tied to a named benchmark in the reported call remarks, and the source material does not provide a workload-specific test supporting it. The comparison also depends on which models are available and how they are scored.
Kimi K3.1 Pro’s weights had not been released by Oct. 26, despite Moonshot’s stated Nov. 4 target. The model’s future access terms and licence were not confirmed in the supplied material. More broadly, price comparisons use listed rates and do not establish total costs across different workloads, service levels or hardware utilization. Actual deployment costs remain company-specific.
The Next Checks for Buyers
The first near-term test is whether Moonshot releases Kimi K3.1 Pro’s weights as promised by Nov. 4 and under terms enterprises can use. Updated Artificial Analysis scores could then show how the downloadable model compares with proprietary systems, though a leaderboard score will not replace task-specific trials.
For organizations weighing self-hosting, the next step is to compare their expected token demand and utilization with the full costs of hardware, staffing and operations. They will also need to assess data-control requirements, technical capacity and each model’s licence. IBM’s argument is that open-weight models can make private AI less expensive; whether that works in practice will depend on those factors and on how quickly models, prices and access terms change.
Key Questions
What did IBM CEO Arvind Krishna say about open-weight models?
On IBM’s Oct. 21, 2026, earnings call, Krishna said open-weight models can cost 10% to 25% as much per token as frontier models while being about 90% as effective. The 90% figure was his characterization, not a result attributed to a named benchmark in the reported remarks.
How close are downloadable models to the leading proprietary model?
In the Artificial Analysis index checked Oct. 26, Xiaomi’s MiMo-V2.6-Pro scored 46.3, compared with 57.6 for Claude Opus 5.5, or about 80% of its score. Index results measure performance on the evaluations used there and do not predict every company’s results.
Is Kimi K3.1 Pro an open-weight model?
Not as of the Oct. 26 comparison. Moonshot had promised to release its weights by Nov. 4, but they had not yet shipped. Its score of 55.1 therefore does not establish that enterprises could download or deploy it at that time.
Is it cheaper for a company to run AI on its own servers?
It can be, especially with high and steady utilization, but the supplied hardware comparison leaves out costs such as staff, cooling and networking. Low-cost hosted models may also be competitive, so the answer depends on workload, usage, operational capacity and data requirements.
Source: ThorstenMeyerAI.com
